About the Project
An AI assistant that bridges the gap between passive viewing and active understanding
Voice + Visual
Fuses audio explanations with slide content
Real-time
Instant answers during live sessions
Smart Q&A
Natural language questions anytime
Citations
References to specific slides
The Problem
Online presentations lack tools for active engagement and understanding
Current Challenges
- Attendees join late and miss context
- Verbal explanations are easily forgotten
- Hard to connect earlier slides to later ones
- AI assistants only transcribe audio
- Visual content is ignored
PARSE Solution
- Fuses verbal + visual content in real-time
- Maintains a live semantic model
- Answers questions in natural language
- Provides citations to specific slides
- Understands formulas, charts, and diagrams
Intuitive Dashboard
Your Rooms
Manage your presentation sessions
CS 342 Lecture
LiveCS 464 Lecture
ScheduledCS 470 Lecture
EndedBeside Zoom, never inside it.
The student keeps the lecture exactly as it is. PARSE opens in its own native window and listens alongside.
How a question finds its slide.
Each slide is split into typed patches. A Turkish question lights up the patches that actually answer it.
Tech Stack
Built with cutting-edge AI and modern web technologies
Development
Watch PARSE in Action
A short walkthrough of the desktop companion sitting beside Zoom and answering Turkish lecture questions.
Reports
Every written deliverable produced for CS 491-492, from the first specification document to the final report.
Project Specification Document
Initial project scope, objectives, stakeholders, and high-level deliverables.
Analysis & Requirements Report
Functional and non-functional requirements, competitor analysis, and user stories.
Detailed Design Report
Architecture, subsystem decomposition, sequence and class diagrams.
Final Report
Implementation, ablation studies, test results, maintenance plan, and user manual.
Project Presentation
The PARSE pitch deck used for the senior-project jury and partner conversations. Thirteen slides covering the problem, the architecture, the demo, the market, and the team.
Team Logbooks
Selected weekly meeting notes across CS 491 (Fall 2025) and CS 492 (Spring 2026).
Pre-kickoff — picking a project worth two semesters
Informal first meeting at the engineering café. Surveyed five candidate project ideas (LLM tutor for Turkish high-school students, AI legal assistant, meeting copilot, music recommender, accessibility translator). Eren had spent the previous summer prototyping a meeting bot using the Zoom SDK and pushed hard for the meeting-copilot direction; demonstrated his existing Zoom + Deepgram integration sketch on his laptop. Anıl liked it because it sat at the intersection of vision and language. Aybars cared about it being shippable on a senior-project timeline. Agreed in principle on the meeting-copilot, with the explicit Turkish-first twist Eren proposed. Action items: each member spends the next week reading prior art and the rubric, and we lock the scope at the next meeting.
Kickoff & scope alignment
First formal meeting at EA-Z01. Anchored the product on a single sentence: a Turkish-first study companion that sits beside Zoom and listens. Walked through the Bilkent senior-project rubric together so everyone agreed on what 'done' looks like by May 2026. Eren came with a one-page diagram of the meeting → bot → transcript → answer data path and walked the team through it. Split roles: Anıl owns the ML pipeline architecture and prompt design, Barış owns ML infrastructure and deployment, Eren owns cloud / Zoom / speech-to-text integration, Aybars owns the backend auth + room management, Bora owns the Next.js dashboard. Set up the GitHub repo, decided on trunk-based development with short-lived feature branches, agreed to meet weekly on Mondays, and Eren took the action item to research Recall.ai vs direct Zoom RTMS for the join path.
Tech stack survey & model shortlists
Anıl, Barış, and Eren reviewed candidate models for each pipeline stage. Document layout: shortlisted DocLayout-YOLO over LayoutLMv3 (lighter, easier deploy). OCR: shortlisted Granite-Docling, PaddleOCR-VL, EasyOCR, GLM-OCR for a comparison sprint. Embeddings: bge-m3, multilingual-e5, OpenCLIP for text; CLIP ViT-L-14, mCLIP variants for visual. Generation: Qwen2.5-VL initially, with Qwen3-VL on the radar. Eren reported back from his Recall.ai research: trial-tier gives us 1,000 free bot minutes per month, more than enough for the senior project. Decided to put every model decision through reproducible notebooks before committing — Anıl took the lead notebook authoring role.
Project Specification Document writing sprint
All-day Saturday writing session for the Project Specification Document. Eren wrote the integration-strategy chapter and the cloud architecture diagram. Aybars wrote the backend service description. Anıl wrote the ML pipeline section. Bora drafted the frontend section. Barış wrote the deployment / infrastructure sections. Discovered three open scope questions while writing: should the desktop companion be a real native app or just a PWA (deferred), should we support recordings playback or only live (decided live-only for v1), and should we cover languages beyond Turkish (no for v1, future work). Submitted the draft for supervisor review on Sunday night.
Architecture freeze for the Specification Document
Long debate on the Zoom integration path. Direct Zoom RTMS gives us low-latency desk-share but requires Marketplace OAuth review and native-thread bridging on the bot. Recall.ai handles all of that as a hosted bot but adds an external dependency and trial-tier costs. Eren led the decision-making — he had implemented both approaches in throwaway prototypes the previous week and walked the team through the latency / setup / failure-mode trade-offs. Decided Recall.ai as the primary path for the senior project (faster bring-up, easier failure model) and kept RTMS as the alternative path. Backend settled on FastAPI + Better Auth + PostgreSQL + pgvector. Eren took ownership of the AWS application host (EC2 + Route 53 + ngrok for OAuth callbacks).
AWS account setup & EC2 provisioning
Eren stood up the AWS environment end to end. Created the team AWS account on the AWS Academy student credit, provisioned a t3.xlarge for the application host (Postgres + bot + backend), set up the VPC + security groups, and registered the parse.app domain through Route 53. Configured ngrok tunnels for the Zoom OAuth callbacks during local development. Aybars worked alongside to wire the backend's .env into the EC2 host so the bot service could reach Recall.ai and Deepgram from the cloud. End of session: live Recall.ai webhooks were reaching the EC2 host through the ngrok tunnel and being acknowledged by the bot. Eren also drafted the runbook for re-provisioning the host from scratch in case of disaster.
First end-to-end Zoom + bot smoke test
Eren got the bot to join a real Zoom meeting via Recall.ai. Audio bridges to Deepgram Nova 3, transcripts surface in the bot's logs within ~1.6 s of speech. Backend issued a JWT and the bot subscribed cleanly. No slide pipeline yet, no UI yet — just the data path. Eren also added structured logging around every Recall.ai webhook and the Deepgram WebSocket so debugging future issues stays tractable. Captured one ten-minute Turkish lecture as a fixture for downstream tests. Aybars patched a CORS issue in the FastAPI middleware that blocked the bot's webhook callbacks from registering. Eren staged the fixture in S3 so all team members can replay it locally without paying Deepgram credits.
Slide ingest first prototype (and first negative result)
Anıl and Barış got a first end-to-end slide-to-pgvector pipeline running on Colab using whole-slide CLIP. Tried it on six real Turkish lecture slides — retrieval was near-random. Anıl traced it to the OpenCLIP text tower being trained only on English captions; Turkish queries fell back to bag-of-tokens behaviour. Eren contributed the slide-change detection upstream of this prototype (perceptual hashing on the Zoom deskshare stream so we only invoke ingest on actual slide transitions, not on every video frame). Agreed the patch-level direction is worth investigating: split each slide into typed regions (title, plain text, figure, table, formula), then route them to language-aware vs vision-aware encoders. Anıl took the ablation-notebook lead.
Deepgram Nova 3 integration deep-dive
Eren spent the week tuning the Deepgram Nova 3 integration. Three issues fixed: (1) noisy partial transcripts being broadcast as if final — added a server-side filter that only forwards lines with is_final=true; (2) Turkish speaker diarization producing inconsistent labels mid-session — switched to a simpler 'Lecturer / Student' two-bucket scheme; (3) WebSocket reconnect after a transient network drop was leaving stale connections — implemented an exponential backoff with a 30-s ceiling. Anıl reviewed the bridge layer and proposed tagging each transcript line with the current slide_id at ingest time so retrieval can correlate later. Eren agreed to ship that in the next sprint.
pgvector schema redesign
Barış proposed one row per YOLO patch with two nullable embedding columns (emb_bge for text patches, emb_mclip for non-text patches). This replaces the original two-table design (separate slide_patches with one embedding type). Cleaner, lets the same retrieval query handle both streams. Aybars adjusted the meeting_id isolation rules so all queries continue to scope to a single meeting. Eren confirmed the transcripts table needs a (meeting_id, slide_id) composite index to power the slide-speech correlation he was about to ship — added to the same migration. INIT_SQL rewritten to be idempotent across the migration so re-deploys don't need manual DBA work.
OCR shoot-out results
Anıl presented compare_ocr_and_text_embeddings.ipynb over 100 Turkish slides across 7 OCR engines (docling + EasyOCR, EasyOCR standalone, Granite-Docling, PaddleOCR-VL, LightOnOCR, GLM-OCR, PP-OCRv5) paired with two text encoders (bge-m3 and DistilBERTurk). For the diagnostic query "11. sorunun cevabı nedir", only GLM-OCR + bge-m3 returned the correct slide at rank 1 (turkce-008, dist 0.4245). DistilBERTurk saturated at dist ~0.0958 across all OCR variants. Eren cross-checked the bge-m3 query latency under the bot's runtime conditions and confirmed it would not block the WebSocket fan-out. Locked in GLM-OCR + bge-m3 as the production combination. Paddle was disqualified separately because its CUDA toolkit conflicts with PyTorch in the same environment.
Sentry observability + reproducible deploys
Barış wired Sentry into every PARSE service (FastAPI backend, FastAPI bot, GPU server, Next.js dashboard, Electron desktop companion). Authored a shared before_send hook that strips Better Auth session cookies, X-Proxy-Secret, transcript text bodies, and slide image bytes from every payload before it leaves the host — only metadata stays. Eren paired on the bot's Sentry integration so Recall.ai webhook payload IDs and Deepgram session IDs are captured via set_context (lets the team reproduce a transcript-fan-out bug from the failing webhook). Documented the deploy process: Docker Compose for application stack, native Python service on the GPU host. Wrote the systemd unit so the GPU server survives reboots.
Analysis & Requirements Report submission prep
All-day Saturday session writing the Analysis & Requirements Report. Anıl wrote the architecture chapter and the patch-level retrieval section. Eren wrote the Recall.ai / Deepgram / Zoom RTMS chapter and the AWS deployment chapter. Bora wrote the frontend and UX chapter. Aybars wrote the backend API contract chapter. Barış wrote the persistence and deployment chapters. Eren also drove the use-case modeling — wrote 14 actor-scenario pairs covering the student joining mid-class, asking a follow-up question, regenerating a stale answer, and the lecturer's privacy-related concerns. Submitted draft Sunday night for supervisor review.
CS 491 wrap, demo day rehearsal
Dress rehearsal for the Fall jury demo: dashboard sign-in, Zoom join via Recall.ai, slide ingest, Turkish Q&A. Caught two regressions live: the bot's public WebSocket fan-out dropped occasional lines under high transcript rate (Eren patched with a queue backed by deque(maxlen=200)); the GPU server's slide-image upload path had an off-by-one on bbox clamping that produced 0-pixel-wide patches on edge cases (Anıl patched). Eren ran the demo himself in front of the supervisor and answered the technical Q&A. Took notes on every clarifying question the supervisor asked, used them later to harden the Implementation chapter.
Detailed Design Report submission + Fall semester close
Final Fall deliverable. Submitted the Detailed Design Report (52 pages including subsystem decomposition, sequence diagrams, deployment diagram, persistent data management, and access control sections). Eren wrote the Hardware/Software Mapping chapter and the Sequence Diagrams for the Recall.ai flow. Anıl wrote the Class Diagram and Data Flow Diagram. Quick retrospective on Fall: shipped the live data path end-to-end, picked the right OCR / embedder combination, deferred the desktop companion to Spring — everyone in agreement that the Spring half is for polish + customer outreach, not new architecture.
Holiday recap & Spring planning
Back from the winter break. Reviewed what each member did over the holidays: Eren wrote a draft of the desktop companion's UX flow and built a Figma prototype, Anıl re-ran the OCR shoot-out on a fresh 100-slide corpus and confirmed GLM-OCR + bge-m3 still wins, Barış set up a staging environment on a second EC2 instance, Aybars added rate-limiting middleware to the backend, Bora redesigned the dashboard's auth pages. Eren walked the team through his Figma — agreed the desktop companion is a four-tab window (Transcript, Q&A, Notes, Profile) with a compact 580x620 native macOS feel. Spring planning: 14 weeks to ship + test + write the Final Report.
Recall.ai paid-tier review & contract
Eren spent the week negotiating with Recall.ai sales for a discounted academic-research tier so we could safely run the test sprint without burning out the free 1,000 minutes. Got a verbal commitment for 5,000 free minutes through the senior-project window. Aybars reviewed the data-processing agreement and confirmed nothing in there forces us to share student transcripts back to Recall. Eren also drafted PARSE's own privacy posture document (per-meeting isolation, before_send Sentry redaction, no third-party training pipeline) so we have something to send to dershane prospects who ask.
Spring kickoff — desktop companion as primary surface
After looking at how students actually use Zoom, decided the in-browser dashboard isn't where the Q&A lives during a real lecture — the student already has Zoom open and doesn't want to alt-tab into a browser tab. Re-prioritised the roadmap: the Electron desktop companion (sitting beside Zoom) becomes the primary student surface. Eren took ownership of the session-control flow (paste Zoom URL → Recall bot join → WebSocket subscribe → live transcript scroll → Q&A composer). Bora's dashboard stays as the account-management entry point. Zoom sidebar app demoted to an optional alternative for users who don't want to install the desktop app. Eren also volunteered to handle macOS code-signing and notarisation once the app reached beta.
Electron desktop companion v0
Eren shipped the first runnable build of the Electron desktop companion. 580x620 native macOS window with hidden-inset titlebar, vibrancy=under-window, three traffic-light buttons. Tab strip: Transcript (active), Q&A, Notes, Profile. Composer at the bottom with placeholder 'Bir soru sor...'. Session model: paste Zoom URL → POST /recall/create-bot → save {bot_id, meeting_id} to localStorage → connect WebSocket. Eren and Bora pair-styled the chrome to feel like a native Mac app rather than 'a website wrapped in Electron'. End of session: the app could join a real Zoom meeting via Recall.ai and stream live transcripts into the panel.
Image-text encoder comparison
Anıl ran compare_patch_embeddings.ipynb on 242 Turkish slides (736 patches) across four CLIP variants. English-only CLIP ViT-L-14 returned essentially random top-K for Turkish queries (confirming the November finding). mCLIP xlm-roberta-large-ViT-H-14 had the best quality but ate ~3 GB VRAM. Turkish-LoRA-CLIP did well on text but underperformed on figures. mCLIP base ViT-B-32 (fp16, ~425 MB) was the only candidate that fit the budget while still ranking the right diagram patches at rank 1. Eren ran the recall metrics on the same corpus and confirmed visual retrieval recall@5 = 0.91. Locked in for production.
Patch-level pipeline production rewrite
Anıl pushed the rewritten gpu-server/server.py. Two-stream retrieval: bge-m3 over text patches deduplicated to top-5 unique slides, mCLIP over visual patches as top-5 raw crops. Disk-cached GLM-OCR under $PROJECT_ROOT/patch_ocr_cache/{slide_id}/{patch_id}.txt so re-uploads of the same slide skip OCR entirely (warm-cache ingest drops from 7.8 s to 1.4 s). 5-letter noise filter on OCR strings before bge-m3 embedding. Detailed Turkish system prompt with explicit rules for MCQ format, LaTeX math, no fabrication. Median Q&A latency on the reference 24 GB GPU: 4.6 s. Eren updated the bot's slide-forwarding code to send the cropped slide image instead of the full screen grab — drops the upload payload by 60%.
Notes panel: BlockNote + KaTeX
Eren added the Notes tab to the desktop companion. Picked BlockNote (block-based editor) over rich-text alternatives because it supports inline equations natively. Wired KaTeX in for math: students type $\int x^2 dx$ and it renders inline. Added a custom DrawingBlock so a student can sketch a diagram next to their notes. Most importantly: an 'Ask with notes' button that sends the current notes content as additional context to the next /ask call. Bora reviewed the styling and matched it to the dashboard's design system. Eren also notarized the build and produced the first .dmg installer for the team to try.
Zoom Marketplace OAuth submission
Eren submitted the Zoom Marketplace app for OAuth review. Required scopes were the minimum set: meeting:read, recording:read (for slide-share frames), user:read (for the participant's display name). Aybars wired the OAuth flow into the backend with the standard authorization-code-with-PKCE pattern; tokens stored hashed in Postgres. Eren wrote the privacy disclosure and the data-handling justification for Zoom's reviewers — five pages explaining exactly what PARSE captures, retains, and processes. Review usually takes 5–10 business days. Recall.ai path remains the primary, so any review delay doesn't block the demo.
First outbound emails to dershanes
Barış sent the first cold pitches to four Ankara-area prospects from the customer-tracker sheet: CBA Akademi (KPSS / ÖABT), ALFA Akademi (KPSS / EKPSS), Ankara Dil Akademisi (YDS / YÖKDİL / IELTS), Aykut Hoca YDS (solo brand, YDS / YÖKDİL). Eren drafted the pitch template (Turkish, three short paragraphs, with a 30-second framing of the problem followed by an offer of a free in-person demo at the end of one of their evening sessions). CBA Akademi replied within 24 hours: forwarded to their decision-makers.
First customer feedback responses
Three more dershane replies came in over the week. Ankara Dil Akademisi: 'Ankara'da olmanız avantaj — bir akşam dersinin sonunda 15 dakikalık bir demo dinleyebiliriz.' (interested in walk-in demo). Aykut Hoca YDS: replied over IG DM, asking to circle back before his August YÖKDİL kursu. ALFA Akademi: asked for a demo video and a pricing sheet. Eren prepped the demo video specifically for these prospects — a 90-second clip on the desktop companion answering Turkish questions over a Yapay Sinir Ağları lecture slide. Barış sent the follow-ups with the video link.
Q&A regenerate + slide-correlated transcripts
Eren shipped slide-correlated transcript storage (every transcript line tagged with the current slide_id at ingest time, indexed by (meeting_id, slide_id) so retrieval can pull only the lines spoken while a matched slide was on screen). This was a multi-week effort — required threading the current_slide_id through the bot's deskshare callback, the Deepgram bridge, and the GPU server's /transcripts endpoint, all without dropping any transcript lines during the propagation. Anıl shipped the regenerate flag on /ask: when true, the most recent qa_history row is dropped and sampling is enabled (do_sample=True, temperature=0.9, top_p=0.92) so the re-answer is genuinely different from the discarded one. Default path (regenerate=false) is byte-identical to the previous deterministic decode.
Desktop companion polish + macOS notarisation
Eren polished the desktop companion for the test sprint. Added: regenerate / edit / copy actions on every Q&A pair, an empty-state sparkle icon for the first-time experience, smooth fade transitions between tabs, optional 'Notes attached' pill that lets the student promote their notes into the next Q&A. Fixed the WebSocket reconnect loop that occasionally fired twice on a network drop. Code-signed the build with the team's Apple Developer ID and submitted to Apple's notarisation service. Bora reviewed the typography one more time. End of session: .dmg installer was downloadable, opened cleanly on a fresh macOS box, no Gatekeeper warnings.
User Manual draft + maintenance plan review
Walked through the User Manual draft together (configuration reference, reverse-proxy Caddyfile, systemd unit, three user workflows, FAQ). Eren wrote the entire 'First-Run Walkthrough (Student)' section since he owns the desktop companion's UX flow. Discussed support channels and SLAs: severity-1 acknowledgement target 4 hours, severity-2 one business day. Decided the in-product Help link in the desktop companion's profile panel will open the team's shared inbox during the senior project and migrate to a real ticketing system if the project continues. Eren also wrote the troubleshooting table at the bottom of the chapter.
Pre-test-sprint dry run
Two-day pre-flight check before the dedicated test sprint. Eren ran every flow end to end on the staging stack — sign-in, join via Recall.ai, slide ingest, ask a question, regenerate, view past session, sign out — and produced a punch list of 9 small bugs (mostly UI polish: a missing focus ring on the composer, a misaligned avatar, a Turkish caption that wrapped one line too tight). Anıl fixed two pipeline-side issues Eren found (a stale OCR cache check that wasn't sensitive to file mtime; the prompt was occasionally producing English answers when the question was English-mixed). All nine bugs filed, sized, and assigned by Monday morning.
Test sprint kicked off
Eren coordinated the start of the dedicated week-long test sprint against the staging stack (one Linux x86_64 application host running the Compose stack, one separate 24 GB GPU host running the patch-level GPU server). 54 test cases on the docket across Authentication, Room Management, Transcription, Slide & Patch Pipeline, Question Answering, Zoom Integration, Bot WebSocket Delivery, LiveKit smoke test (legacy), GPU Server / Reset / LaTeX, Desktop Companion, Performance, Security, Reliability, Compatibility, Usability, and Maintainability. Eren personally executed every test card and logged the Created-By / Tested-By / Date-Tested / Result rows.
Test sprint complete
Final tally: 51 strict passes, 2 pass-with-notes (multi-participant video occasionally drops the second remote video on Safari 16 for ~1 s; dashboard render with 50 rooms takes 1.4 s vs the 1 s soft target — both filed as low-severity), 1 hard failure (TC-REL-003: Safari WebRTC reconnection on network drops >30 s, scoped only to the legacy self-hosted LiveKit path; the primary Recall.ai path is unaffected because its WebSocket fan-out auto-reconnects within ~1.4 s). Strict pass rate 94.4 %. Filed as ISSUE-217 for v1.1. Eren wrote the Test Results Summary section of the Final Report based on his sprint notes.
Final Report draft review
Full-pass review of the Final Report. Caught: a few residual references to the old whole-slide CLIP architecture in the Subsystem Services chapter (Anıl patched), the Maintenance plan was missing the Sentry observability section (Barış added), and the Teamwork chapter needed the Meeting Objectives milestone-lead column (added). Reviewed the User Manual chapter line by line against the actual ENGINEERING.md / README in the repo. Eren reviewed the Test Cases chapter end to end and tightened the Result strings on five cards. Locked the chapter structure.
Final Report freeze
All chapters reviewed by at least two team members. Locked title-page metadata: Supervisor Ayşegül Dündar Boral, Innovation Expert Sezai Artun Özyeğin, Course Instructors İlker Burak Kurt & Mert Bıçakçı. Submitted the Final Report PDF and the 60-second demo video to the capstone portal. Eren produced the demo video — a 90-second cut showing the desktop companion sitting beside Zoom during a Turkish lecture, the composer typing in a Turkish question, the streaming answer, and the regenerate flow. Wrote the team retrospective: what worked (notebook-driven model decisions, rotating leadership on big workstreams, Eren owning the end-to-end cloud + Zoom + STT integration), what to do differently next time (start outreach earlier, dedicate test sprint earlier).
Capstone fair preparation
Capstone fair is in five days. Eren built the A1 poster in Figma (PARSE wordmark, three-column architecture diagram, four stat tiles, team avatars, QR code linking to parse.app). Printed two copies — one for the booth, one as backup. Anıl rehearsed the technical walkthrough; Eren rehearsed the demo and the customer-traction story. Built a dedicated demo-day staging environment with a pre-loaded session so the bot doesn't have to cold-start in front of the jury. Wrote a one-page elevator pitch card to hand to interested visitors.
Capstone fair & demo
Live demo at the Bilkent CS senior-project fair. Eren ran the demo: paste Zoom URL → bot joins → live Turkish transcript → ask 'Yapay sinir ağı nedir?' → grounded answer in < 5 s with citation to the slide. Demonstrated the desktop companion sitting beside Zoom during a recorded Turkish lecture, ran four pre-rehearsed questions through the patch-level pipeline, walked the jury through the architecture poster. Audience questions about the OCR shoot-out methodology and the regenerate feature. Closed by linking parse.app and the GitHub repo. Final reflection: the project is ready to continue past the senior-project window if any of the dershane prospects converts to a pilot.
Weekly team meetings; per-member personal logbooks are maintained in the team's internal Notion workspace.
Customer Conversations
Four early conversations with Turkish exam-prep dershanes. Outreach started in March 2026 — these are the prospects who replied.
CBA Akademi
Forwarded internallyDershane · KPSS / ÖABT / AGS / ALES / DGS
Email · 2026-03-13
Merhabalar Barış Bey, mesajınızı yetkili birimimize ilettik. İlgili ekiple görüşme sonrası size dönüş yapılacaktır.
First positive reply received within 24 hours. Pitch was forwarded to their decision-makers. Follow-up scheduled for late April after the team's internal review.
Ankara Dil Akademisi
Demo interestDershane · YDS / YÖKDİL / IELTS / TOEFL
Email + walk-in offer · 2026-03-28
Ankara'da olmanız avantaj — bir akşam dersinin sonunda 15 dakikalık bir demo dinleyebiliriz. Hocalarımızla planlayıp size geri döneriz.
Open to a short in-person demo at the end of one of their evening live sessions. They wanted us to confirm whether transcripts are stored on-premise or in the cloud before agreeing on a date.
Aykut Hoca YDS
Interested, slow channelSolo-instructor dershane · YDS / YÖKDİL
Instagram DM · 2026-03-30
İlginç görünüyor. Şu an yoğun sezon, Ağustos kursundan önce bir telefon görüşmesi yapabiliriz.
Solo brand, fast initial reply over IG DM. Said the live Q&A loop in his Tuesday/Thursday/Friday evening classes is exactly the kind of thing PARSE addresses. Asked to circle back closer to the next term.
ALFA Akademi
Needs more informationDershane · KPSS / EKPSS / DGS / ALES
Email · 2026-03-28
AI destekli araçları araştırıyoruz; lütfen demo videosu ve fiyatlandırma bilgisi gönderir misiniz?
Cautious but engaged. Replied asking for a demo video and a pricing sheet. Indicated they are already evaluating AI tooling — would want to compare apples to apples before committing to a pilot.
These conversations focus on Turkish dershanes only. Broader prospect tracking (medical-prep, EdTech bootcamps) is maintained in the internal customer-tracker sheet and is not included here.
Team
Anıl Kılıç
Aybars Buğra Aksoy
Bora Yetkin
Eren Berk Eraslan
Barış Yaycı