Currently building — open to collaborations

I build systems that retrieve,
coordinate, and verify.

I'm Aditya. I work on the infrastructure layer of intelligent systems — agent runtimes, retrieval and memory, verifiable execution, and the quieter pieces that turn AI demos into production infrastructure.

Explore systems Read field notesRésumé
Selected systems

A few things I've built, and one I'm still building.

Most of my work lives in the seams between agents, retrieval, and verification. Three of those projects are below; the rest are quieter, but the index follows.

Memora

Verifiable long-term memory infrastructure for autonomous AI agents.

Write path: the SDK AES-256-GCM encrypts the payload, pins ciphertext to IPFS via Pinata, submits a consensus-ordered sequence number to Hedera HCS, calls commitMemory() on a Hedera EVM registry contract, and registers the AES key with a challenge/response key broker — all in a single indexer write.

Read path: query the Supabase index for the memory reference, fetch ciphertext from IPFS, authenticate with the key broker via a signed challenge (key release gated by on-chain permission check), decrypt, then re-hash to verify payload integrity. Supabase is a rebuildable cache — the authoritative record is on-chain.

StackTypeScript · Node.js · Hedera HCS · Solidity · IPFS · Supabase · Express · pnpm workspaces
GitHub Case study
fig. 01 — write / read pathmemora
fig. 02 — trust boundary analysis

TrustGraph

Deterministic trust-boundary vulnerability analyser for Solidity contracts.

Applies a four-predicate vulnerability predicate — external visibility, unverified payload parameter, critical state mutation, absent caller guard — to flag CrossCurve-style exploits. An 8-node LangGraph pipeline loads contracts, scores each function, classifies trust assumptions, generates executable Foundry PoC tests, and emits Markdown and JSON reports. Gemini 2.5 Flash optionally enriches explanations but never overrides deterministic severity.

StackPython · LangGraph · Pydantic · Foundry · Solidity · Gemini 2.5 Flash · Typer · VS Code
GitHub
fig. 03 — runtime topology

AgentOS

Local execution infrastructure for coding agents. Prototype.

Isolated git worktrees, controlled command execution, workspace-level guardrails, command allowlists, and diff inspection. Runtime tooling for bounded autonomous development workflows — actively building.

StackTypeScript · Node.js · Git · Shell
GitHub
Also building
Retrieval Adaptation Pipeline
Built a 6,044-example YouCook2 disruption benchmark and VLM evaluation pipeline for missing-ingredient procedural adaptation. Tested text, vision, retrieval, and reranking across Qwen/Gemma models; retrieval dominated at 2B, while stronger VLMs shifted toward vision-grounded retrieval.
research
SmartInvoiceExtractor
Automated medical receipt and invoice parsing pipeline for handwritten and printed hospital documents. Regex + Gemini 2.5 Flash fallback; ≈98.2% field-level accuracy, ≈65% inference cost reduction, ≈80% manual correction reduction.
shipped
Sageo / MOI Labs
Identity and interaction verification platform for AI agents on MOI Chain. Backend REST APIs, agent authentication, hashed interaction proofs, and encrypted off-chain logs with on-chain hash commitments.
collaboration
Clinical Speech ML Pipeline
End-to-end speech preprocessing and federated training pipeline for clinical audio. Whisper + Librosa feature extraction, 200+ samples, Dockerized ingestion, 1.5× throughput, 15% accuracy gain under distribution shift.
archived
Experience

Where this thinking came from.

Jul 2025 — Aug 2025Solsten
Software Engineering Intern

Medical invoice parsing pipeline with hybrid extraction and LLM fallback.

Built an automated parsing pipeline for handwritten and printed medical receipts and invoices. Combined regex-based extraction with Gemini 2.5 Flash for missing or low-confidence fields, reducing inference costs by ≈65%. Refined extraction logic with DSPy evaluation and pytest validation, reaching ≈98.2% field-level accuracy and reducing manual correction workload by ≈80%.

Python · Vertex AI · Gemini 2.5 Flash · DSPy · pytest
Sep 2025 — Dec 2025MOI Labs × BoilerBlockchain
Software Engineer / Project Lead

Agent identity and interaction verification on MOI Chain.

Built Sageo, an identity and interaction verification platform for AI agents on the MOI chain. Implemented backend REST APIs, agent authentication logic, hashed interaction proofs, and a hybrid logging system with encrypted off-chain payloads and on-chain hash commitments for verifiable audit trails.

TypeScript · React · Node.js · MOI Chain · Google A2A Protocol · Docker · CI
Feb 2025 — Aug 2025Kong Labs
Undergraduate Research Assistant

Clinical speech preprocessing and federated training pipeline.

Built an end-to-end clinical speech ML pipeline for scalable preprocessing and federated training across distributed hospital datasets. Implemented low-latency feature extraction with Whisper and Librosa, processed 200+ clinical audio samples, Dockerized ingestion and preprocessing, improved dataset throughput by 1.5×, and designed federated workflows that improved classification accuracy by 15% under distribution shift.

Python · PyTorch · Whisper · Librosa · Docker · Federated Learning
Jan 2025 — presentBoilerBlockchain, Purdue
Technical Course Lead

10-week blockchain engineering curriculum — 100+ students certified.

Designed and delivered a 10-week blockchain engineering curriculum covering EVM, Solidity, ERC standards, DAOs, DEX mechanics, and L2 scaling. Taught and certified 100+ students, improved course completion rate by 15%, and expanded the active developer team by 3×.

Solidity · Foundry · EVM · TypeScript
Field notes

A private notebook, accidentally public.

Working theories, half-formed opinions, things I keep relearning. Mostly about infrastructure for systems that try to think. Read at your own risk.

2026 · 05 · 09infra

Most AI infra problems are observability problems wearing a costume. If you can't see what the agent did, you can't fix the prompt — you're just guessing.

2026 · 04 · 22agents

Autonomous systems need memory more than intelligence. A 7B model with disciplined recall beats a 70B with goldfish syndrome.

2026 · 03 · 30systems

Deterministic systems age better than clever prompts. Taste, in design, is which parts you refuse to leave to the model.

2026 · 03 · 11verification

Verification is what turns agent workflows from demos into infrastructure. Without it, every output is a vibe; with it, the same output is evidence.

2026 · 02 · 28retrieval

Retrieval is not always better. In the 2B VLM setting, retrieved procedural analogies did most of the work; at stronger scales, visual grounding started to matter more. The retrieval strategy was a function of model capacity, not a universal rule.

2026 · 01 · 19runtime

"Give the agent more tools" usually means "we don't have a planner." More tools without a planner is just a bigger combinator space to get lost in.

Currently investigating

Open questions I keep returning to.

Some of these turn into systems. Some turn into one-page writeups. Some just stay open until they don't.

When does forgetting help?

Designing TTL and decay policies for episodic memory. Most of the interesting behavior is in what an agent chooses to drop, not what it stores.

active
microVMs vs. seccomp for coding agents

Measuring blast radius under hostile tool use. The interesting question is what bounded autonomy actually costs you in latency and ergonomics.

active
Retrieval evaluation beyond MRR

Semantic evaluation that survives paraphrase and adversarial rewrite. The current benchmarks are mostly testing whether you remember the answer, not whether you can find it.

writing
Compact attestation formats

Merkle-DAG transcripts versus zk-friendly traces. Verifiers want small; agents want expressive. The interesting work is where they meet.

active
Local-first autonomous workflows

Running planner, retrieval, and tools on-device. What breaks first at 8GB RAM, and which parts should never have been in the cloud anyway?

parked
Trust boundaries in AI systems

Who signs what, when, and what does the verifier actually verify? Most "trust" claims in AI are vibes-with-a-checkmark.

writing
Stack

What I actually reach for.

Grouped by function, not by category. Not a checklist — the things below are the ones I've shipped with, broken, and would pick again.

Backend & infra
Python·TypeScript·Node.js·FastAPI·Postgres·Redis·Docker·Supabase
AI systems
PyTorch·vLLM·DSPy·Hugging Face·BGE·sentence-transformers·Vertex AI
Runtime & tooling
OpenTelemetry·pytest·Foundry·Git·Docker·CI/CD·neovim
Verification & chain
Solidity·Foundry·EVM·Hedera HCS·IPFS·MOI Chain
Frontend & product
React·Next.js·Tailwind·TypeScript·Figma
Habits
writing design docs·eval-first·small commits·read the source·observability before features
Get in touch

Building something load-bearing?

I'm open to backend SWE, AI infrastructure, and platform engineering roles — and to ambitious technical collaborations on agents, retrieval, runtime, or verifiable systems. If you're working on something that needs to survive past the demo stage, send me a note.

Fastest channel · email or LinkedIn DM. Replies usually under 24h.
emailakuniyil@purdue.edugithubgithub.com/Akk525linkedinin/aditya-kuniyil-kattilrésuméadityakk.com/resume.pdf