Talha Khalil

Event-driven audio and video to bilingual English/French subtitles on AWS.

• Built a serverless pipeline where an S3 upload triggers Lambda validation (magic bytes plus ffprobe), Amazon Transcribe with language identification, and an EventBridge-driven Amazon Translate step that keeps translated .srt/.vtt cue counts and timestamps identical to the source; a 14.4 s clip goes from upload to subtitles in about 8 s for roughly $0.009 per run.
• Added an arm64 container Lambda on ECR with a pinned static ffmpeg to extract audio from .mov, .mkv and .avi, plus idempotent job names, DLQs with retries, and S3-as-database outputs with no separate datastore.
• Shipped a Cognito-protected React player behind CloudFront (CSP, HSTS) and a JWT-authorized API Gateway + FastAPI Lambda, all least-privilege IAM, deployed and torn down with one Terraform command at about $1.1 per month idle.

Final year project: AI-driven virtual internships for students, with mentor review.

• Built skill assessment across 5 domains where a scikit-learn LogisticRegression classifier turns per-domain MCQ accuracy into a recommended track, then matched students to a 60-task bank with per-domain FAISS indices (all-MiniLM-L6-v2 embeddings) re-ranked by a local Ollama LLM, falling back to FAISS order on malformed output.
• Graded every submission with a JSON-constrained LLM rubric (score bands, feedback, 3–5 criteria) and flagged plagiarism through a per-task FAISS index written incrementally at a 0.9 cosine-similarity threshold, feeding a mentor desk with a flag-first queue, side-by-side match review and score overrides.
• Shipped a token-streamed career-guidance chatbot over SSE grounded in each student's scores and open tasks, plus progress reports and an auto-generated portfolio with achievements, a revocable public share link and PDF export, on a Django + Inertia.js + React monolith backed by MongoDB.

RAG exam generation grounded in source textbooks.

• Built a RAG pipeline grounding every generated exam question in the educator's own uploaded textbook, using recursive character chunking (1000 chars / 200 overlap), Gemini embedding-001 vectors in ChromaDB, and top-k=5 similarity retrieval to produce a 40-question paper in 90 seconds at roughly $0.12 per paper.
• Designed Bloom's Taxonomy-mapped prompt templates giving teachers control over difficulty tier and question type (MCQ, short, long), carrying retrieval metadata (page, chapter) through so every question cites its source.

Multi-agent ingestion and ranking pipeline.

• Built a 5-stage daily ingestion and LLM pipeline across 3 sources (YouTube transcripts, OpenAI and Anthropic RSS), processing roughly 25 items per day with dedup-on-insert and per-stage failure isolation so one failing source never halts the run.
• Implemented a two-agent LLM layer with model-tiered routing for cost, pairing a digest agent (gpt-4o-mini) that summarizes each item with a curator agent (gpt-4.1) that scores 0–10 relevance against a structured user profile via Pydantic structured outputs, running the full pipeline at roughly $5 per month.