Onkar Jadhav

Final-year Computer Science student at PCCOE (9.08 CGPA) with hands-on experience building end-to-end AI systems, including an open-source LLM governance benchmark evaluated across the US, India, and EU legal scenarios, and a deep learning equity prediction system covering Nifty100 NSE equities. Proficient in Python, React.js, Node.js, FastAPI, and LangGraph.

Bachelor of Technology in Computer Science

Pimpri Chinchwad College of Engineering, Savitribai Phule Pune University
Nov 2022 -- Jun 2026
CGPA: 9.08 (up to 7th semester)

Relevant Coursework

• Machine Learning
• Data Structures & Algorithms
• Database Management Systems
• Operating Systems
• Computer Networks
• Object-Oriented Programming

Higher Secondary Certificate

Shri Mhalsakant Junior College, Maharashtra State Board
Jun 2022
Score: 89.17%

StockXpert — AI-Powered Equity Prediction & Stock Recommendation Platform

Tech Stack: Python, TensorFlow, Next.js, FastAPI, TypeScript, Tailwind CSS, Cloudflare R2, AWS
Duration: Nov 2025 -- Present

• Architected a hierarchical deep learning system combining ResNet, GRU, Bi-LSTM, and multi-head attention to capture cross-horizon price dependencies across Nifty100 NSE equities, outperforming single-architecture baselines.
• Constructed a leakage-free feature pipeline from raw OHLCV data incorporating RSI, MACD, ADX, Bollinger Bands, SMA/EMA distance, and volatility indicators across strict train/validation/test splits.
• Implemented a multi-objective Huber loss with directional regularization and tail-event oversampling, achieving a 12% improvement in short-horizon return correlation over baseline and reducing directional error on high-volatility equities.
• Designed a snapshot-first serving layer with Cloudflare R2 storage and market-aware TTL caching, delivering sub-second API responses across all 100 Nifty stocks during market hours while eliminating redundant inference costs.
• Built a full-stack trading dashboard (Next.js + FastAPI) featuring a live stock screener with real-time alerts, OHLCV candlestick charts, technical indicator overlays, and forecast tables (1D–10D) with calibrated confidence scores.

GovBench — Open-Source LLM Governance Benchmark

Tech Stack: Python, LLM Evaluation, AI Safety, Sentiment Analysis, Hallucination Detection
Duration: Mar 2026 -- Present

• Designed and open-sourced GovBench on GitHub, an LLM evaluation framework assessing model readiness for judicial and governmental deployment across 6 pillars: demographic bias, procedural integrity, corruption resistance, jurisdictional awareness, transparency, and minority protection.
• Implemented controlled demographic isolation testing across 12 identity variants per scenario with naturalistic prompting, measuring inherent bias in bail, sentencing, welfare, and immigration decisions across the US, India, and EU legal systems.
• Evaluated 4 production LLMs (Claude Sonnet 4.6, Gemini Flash Lite, DeepSeek V4, GPT-OSS 120B) across baseline, pressure, and adversarial modes, revealing top performers scored 90%+ overall while identifying failures in minority protection (GPT-OSS: 63%) and jurisdictional awareness (DeepSeek V4: 56.7%).
• Built an automated scoring pipeline combining sentiment variance analysis, hallucination detection, and position drift tracking to produce composite deployment-readiness grades across evaluated models.

PromptForge — Agentic Prompt Engineering Platform

Tech Stack: Node.js, Express.js, BullMQ, Convex, OpenRouter, React.js
Duration: Jan 2026 -- Feb 2026

• Developed a platform for designing, testing, and optimizing LLM prompts, enabling developers and AI product teams to generate, refine, and structure prompts for faster delivery and more reliable outputs.
• Architected a creator-critic-evaluator pipeline that iteratively improves prompts through automated feedback, scoring clarity, completeness, constraints, and token efficiency across multiple versions.
• Implemented cross-model routing and benchmarking across OpenAI, Anthropic, Gemini, and open models, enabling direct comparison of quality, latency, and cost per prompt for model selection decisions.
• Built a fault-tolerant async backend using BullMQ workers and Convex persistence, supporting idempotent job retries and end-to-end logging for reliable prompt regression testing at scale.