Projects
RecruitPilot
End-to-end AI recruitment platform combining resume screening, AI-driven interviews, and real-time proctoring.
- Candidate pipeline board tracking applicants from applied through phone screen and technical interview stages
- Live video interview panel with real-time proctoring alerts for lost window focus and background voice detection
- Auto-scores candidates on technical skills and communication, with editable hiring-manager notes per session
Slate
Intelligence engine for investigative journalism automating deep-dive research, narrative structuring, and editorial auditing.
- Research board linking leaked documents, filings, and whistleblower sources into a single entity map and timeline
- Flags contradictions between sources and tracks citation coverage percentage on unverified claims
- Auto-builds a narrative outline and a risk-level report (high/medium) for each investigation section
SiteBOT
Production-ready agentic chatbot platform featuring ultra-low latency multilingual voice mode and dynamic UI streaming.
- Routes English/Hindi voice conversations through Sarvam AI speech recognition into Groq-hosted Llama 3
- Renders dynamic UI cards (order status, retry actions) directly inside the chat based on model output
- Averages 382ms response latency, with a live voice waveform monitor for support agents
Cricket Video Analyzer
Real-time cricket match analysis utilizing Ultron Live SDK for visual commentary and Whisper for audio processing.
- Tracks ball trajectory and classifies shot type (cover drive, square cut) from live video with 96%+ confidence
- Whisper-based commentary transcription paired with auto-generated visual commentary per ball
- Live wagon-wheel breakdown and over-by-over run-rate projections update as the match progresses
Shared Ride Engine
High-performance Redis-backed matching and pricing engine for shared ride pooling and solo assignments.
- Real-time dispatch map matches riders on overlapping routes using Redis geospatial lookups at ~34ms latency
- Scores each potential pool match and estimates solo vs. shared fare before confirming a route
- Live driver availability, route detour %, and a full pricing breakdown surfaced to dispatchers
Mental Health AI
Local symptom classifier and chatbot fine-tuned on Qwen using Apple Silicon. Served via FastAPI.
- 14-class symptom classifier (anxiety, low mood, stress/burnout, and more) with a confidence score per session
- Runs entirely locally on Apple Silicon through a fine-tuned Qwen2.5 model — no user data ever leaves the device
- Generates a risk-level assessment and recommended next steps at the end of each triage conversation
Multi-Agent Workflow Platform
Production system routing user requests to specialized AI agents for emails, documents, and workflows.
- AgentBrain Pipeline creates specialized agents through autonomous research and context fusion instead of static prompting
- Routes 1,000+ daily tasks across email, documents, calendars, spreadsheets, and voice systems in live enterprise use
- Unified multi-LLM architecture built with FastAPI, Redis, and a Next.js control dashboard
NL-to-SQL Media Monitoring Chatbot
Plain-English chat interface for a 200K+ publication database with Redis-backed SQL execution.
- Translates plain-English questions into SQL over a 3.6GB MySQL database of 200K+ press articles
- LangChain handles query generation; GPT-4 formats and summarizes the returned results
- Redis caching on repeated queries cut average response time from 5s down to 2s
Delhi PM2.5 AQI Forecasting
Neural network built from scratch in NumPy to forecast Delhi PM2.5 levels using six years of data.
- Neural network written entirely from scratch in NumPy — no ML frameworks — including a custom backprop implementation
- Trained on six years of real CPCB sensor data, reaching R² = 0.9501, ahead of a Random Forest baseline
- Final-year project pairing classical feature engineering with the from-scratch model
Semantic Memory Engine
Semantic memory engine for local LLMs using Symbolic Prompt Language (SPL) for 10x prompt compression.
- Converts long conversations into Symbolic Prompt Language (SPL), a compressed representation the model reads directly
- Achieves roughly 10x prompt compression compared to passing raw conversation history
- Qwen2.5 fine-tuned via QLoRA to reason from compressed SPL tokens without decompressing back to words, served locally through Ollama