Jan 2025 – Present · 1 yr 2 mos

• 1. Designed and deployed a self-improving multi-agent dialogue system (DSPy, gRPC, internet search) with:
• 🤖 Real-time human–AI group conversations among emotional, professional, and domain-specific agents.
• 🧠 Continuous learning through intent and sentiment analysis to dynamically update agent personas.
• 🌐 Internet and vector-store integration for up-to-date, personalised, and context-aware responses.
• ⚙️ A global–local prompt optimisation pipeline that adapts each user’s persona while improving overall system intelligence.
• 2. Fine-tuned LLMs and encoder-only models (BERT, ModernBERT) for classification tasks, achieving up to 92% F1 accuracy on benchmark datasets.
• 3. Built automated prompt optimisation pipelines, improving model performance by 10–12% and reducing inference costs by 85%.
• 4. Optimised and deployed LLMs for Android (GGUF) and iOS (q-safetensors), cutting memory footprint by 50% while maintaining comparable accuracy.
• 5. Leveraged Unsloth + TRL with GRPO-style training (Deepseek) on Gemma 3, achieving 18% accuracy gain and edge-compatible fine-tuned variants.

Oct 2023 – Dec 2024 · 1 yr 2 mos

• 1. Built a real-time market insights agent (DSPy, LLMs, web scraping) that processed 100% of NSE announcements within 2–5 mins, improving precision to 98% vs. 1–9 hours on traditional systems.
• 2. Implemented STORM, the world’s first deep-research financial RAG system, reducing report creation time by 90% and increasing accuracy by 70%.
• 3. Developed ColPali/ColQwen Image-RAG pipelines, enabling PDF-based financial Q&A without OCR — achieving 90% higher accuracy in edge cases.
• 4. Engineered Neo4j Knowledge Graphs (100k+ nodes) integrated with DSPy for cross-domain data connectivity.
• 5. Created Text2Cypher NL Query System (Neo4j, DSPy, Meilisearch), achieving 95% query accuracy in production.
• 6. Designed hybrid retrieval pipelines (Cohere, HyDE, BM25, DSPy) across FAISS, Pinecone, and Weaviate for state-of-the-art retrieval and ranking.
• 7. Integrated MCP + ReAct for advanced agentic reasoning in finance-focused AI agents.
• 8. Built autonomous web agents for verified information aggregation with authenticity scoring.
• 9. Developed an AI Podcast System using Whisper, ElevenLabs, and Pinecone for retrieval-based dynamic content generation.

Jan 2023 – Oct 2023 · 9 mos

• 1. Achieved 95% personalised image accuracy with 70% less training data using Dreambooth and custom face cropping pipelines.
• 2. Deployed multi-platform diffusion pipelines (AWS, GCP, Lambda Labs, Replicate), ensuring 99.9% uptime.
• 3. Integrated IP Adapter for efficient similarity-based inference, cutting processing time by 85% and costs by 70%.
• 4. Implemented inSwapper128.onnx-based face swapping with 98% realism success rate.
• 5. Improved image inference pipeline efficiency by 60% using HuggingFace + Kubernetes architecture.

Mar 2022 – Dec 2022 · 9 mos

• 1. Designed KYC and authentication services (gRPC, OAuth, JWT, S3), handling discrepancies between multiple identity sources such as Aadhaar and PAN, and later integrated DigiLocker for enhanced security and reliability.
• 2. Built real-time analytics pipelines (Kinesis, Lambda, Athena, Glue) for large-scale game data.
• 3. Developed a secure 2FA system, integrated gRPC and REST services, and implemented a referral system to enhance user engagement and security.
• 4. Developed analytics dashboards with QuickSight to visualise user behaviour, improving analysis speed by 60%.
• 5. Built backend API documentation to improve developer onboarding and maintainability.

Oct 2021 – Mar 2022 · 5 mos

• 1. Built and fine-tuned DCGAN models to enhance profile picture (PFP) generation quality.
• 2. Conducted early LLM research (GPT-3, GPT-J, GPT-Neo) for chatbots and dynamic NFT generation use cases.
• 3. Trained ML Agents in Unity using reinforcement learning and explored potential applications as enemies in third-person shooter games.
• 4. Experimented with NeRF technology to generate 3D voxel models from image datasets.