🎓 Case Western → Cornell Tech | 📄 Solo ICLR 2026 researcher | 🛠️ Founder-researcher-engineer
I build applied AI systems that leave the demo and touch messy real workflows: operators, buildings, cameras, documents, data pipelines, and the tools around them.
- 📄 Beyond Vector Search - solo ICLR 2026 workshop research on hallucination-free financial reasoning with chunk-centric knowledge graphs.
- 🎙️ Agent Voice - always-on local voice front-end for Hermes: wake word, local STT, streaming TTS, and multi-turn spoken tool use.
- 📷 HermesCamera - camera control for Hermes Agent: photo capture, timed video, pan/tilt/zoom movement, room scans, and AI interpretation through
agy. - 🧠 agent-core - TypeScript agent infrastructure experiments.
- 🎬 ReelAutomation - autonomous reel generation, from idea to deployment.
- ✍️ rithvik.journal - commentary on AI, building in public, and Instagram/TikTok reel experiments.
I'm quietly building in the property management operations space. It is still a work in progress, so I am keeping the product details private for now.
I was the solo researcher behind Beyond Vector Search: Hallucination-Free Financial Reasoning with Chunk-Centric Knowledge Graphs, presented as a poster at the ICLR 2026 Workshop on Advances in Financial AI.
The work explores a graph-grounded alternative to pure vector retrieval for financial documents, with a retrieval path that can be followed and audited: entity → relationship → chunk.
I keep coming back to systems where AI has to interact with the world:
- Cameras, sensors, and robotics-adjacent control loops
- Agent tools that can inspect state, take action, and recover from failures
- Data pipelines that turn messy operational inputs into useful decisions
- Developer tooling for faster prototyping, evaluation, and iteration
- 🎓 Cornell Tech - Master of Engineering in Computer Science
- 🎓 Case Western Reserve University - undergraduate degree
- 🧪 Research - solo ICLR 2026 workshop poster
- 🛠️ Focus - applied AI, ML systems, automation, and full-stack product engineering
| Project | What it explores |
|---|---|
| Agent Voice | Always-on local voice interface for Hermes with wake-word detection, progressive local STT endpointing, streaming TTS, and multi-turn conversation state. |
| HermesCamera | Python library and Hermes Agent plugin for controlling a local UVC camera, capturing media, moving pan/tilt/zoom, and interpreting output with agy. |
| agent-core | Minimal TypeScript runtime core for deterministic agent turn orchestration, event streams, session state, and pluggable execution hooks. |
| ReelAutomation | Autonomous LLM workflow for planning, creating, and posting Instagram reels. |
| RetinaLiteNetClassifier | Deep learning classifier built around a recreated RetinaLiteNet encoder and MHA bottleneck. |
| InceptionV4_parallelization | Python parallelization work around InceptionV4. |
| Area | Tools and Patterns |
|---|---|
| LLM systems | Agents, RAG, chunk-centric knowledge graphs, GraphRAG, evaluation, tool use |
| Machine learning | PyTorch, TensorFlow, computer vision, adversarial training, optimization |
| Backend and infra | Python, FastAPI, Node.js, Docker, async pipelines, FAISS, Neo4j, AWS |
| Product engineering | TypeScript, React, APIs, dashboards, automation, operational workflows |
| Systems and hardware | Linux, local voice agents, cameras, robotics-adjacent control loops, scripts, device automation |
- Build software that connects intelligence to action.
- Turn vague, high-leverage workflows into working products.
- Make AI systems measurable, inspectable, and useful.
- Ship prototypes quickly enough that reality can correct the idea.
- Stay close to users, operations, and the real failure modes.
