Featured systemActive
Memora Sentinel
Verifiable execution provenance infrastructure for autonomous systems.
I build production systems — backend infrastructure, blockchain protocols, and AI tooling.
Purdue CS student building across smart contract systems, AI infrastructure, and backend engineering — retrieval pipelines, blockchain protocols, agent runtimes, and production tooling.
Current focus
- smart contract systems
- AI infrastructure
- backend engineering
- blockchain protocols
- production tooling
Latest note5 min read
SmartInvoiceExtractor
Regex extracts high-confidence fields from pdfplumber-parsed PDFs; Gemini fills gaps when critical fields are absent. A quality gate, JSON normalisation layer, and regex suggestion loop keep the LLM call rate low and the patterns improving. Production metrics are from the proprietary Solsten deployment — the public repo is an earlier local prototype.
Read noteCurrently Available
Open to backend SWE, blockchain engineering, AI infrastructure, and platform engineering opportunities.
Let's build systems that last.
Get in touchExhibits 01–04Selected systems
Open the archiveMemora Sentinel
Verifiable execution provenance for autonomous agents — what executed, who authorised it, and can the history be independently verified.
TrustGraph
Deterministic trust-boundary analyser for Solidity contracts.
SmartInvoiceExtractor
Regex-first PDF extraction pipeline for semi-structured medical invoices.
98.2% field accuracy · 65% lower inference cost · 80% manual correction reduction
Python · pdfplumber · Gemini 2.5 Flash · Vertex AI · DSPy · pytestOpen file
AgentOS
Local-first operating environment for supervising and steering coding agents.
All builds10 entries on record
Core projects
TrustGraph
ShippedDeterministic trust-boundary analyser for Solidity contracts.
Python · LangGraph · Foundry · Solidity · Gemini · VS Code extension
WebCoder
Shipped · MVP completeAI-native build platform for Web3 apps — generate, audit, simulate, and deploy from natural language.
Next.js 14 · LangGraph · Foundry · Solidity · Anvil · Slither · wagmi v2 · Claude Sonnet · Vercel · pnpm workspaces
CHITTA
Shipped · deployedAI-assisted wind site screening for terrain-aware renewable energy planning.
Next.js · FastAPI · Python · LangGraph · PostgreSQL · NASA POWER · Mapbox GL · ReportLab · Docker · Railway
Improvisational Reasoning in VLMs
Research · Purdue CS490 · co-authoredResearch paper: when do VLMs benefit from visual grounding vs. retrieval for procedural adaptation?
Python · PyTorch · Qwen2-VL · Gemma · vLLM · FAISS · BGE embeddings · sentence-transformers · Hugging Face
Mano
Co-founded · LiveCo-founded AI deal coordinator for CRE acquisitions — live at mano.network.
Next.js · TypeScript · Supabase · OpenAI · Resend · Tailwind
AgentOS
Active · v2.0 AlphaLocal-first operating environment for supervising and steering coding agents.
React 19 · TypeScript · Vite · Tailwind · Framer Motion · Tauri · Rust · Ollama
Applied work
SmartInvoiceExtractor
Shipped · Solsten 2025Regex-first PDF extraction pipeline for semi-structured medical invoices.
98.2% field accuracy · 65% lower inference cost · 80% manual correction reduction
Python · pdfplumber · Gemini 2.5 Flash · Vertex AI · DSPy · pytest
Clinical Speech ML Pipeline
Archived · Kong Labs 2025End-to-end speech analysis and federated training pipeline for clinical audio.
200+ samples · 1.5× throughput · 15% accuracy gain under distribution shift
Python · PyTorch · Whisper · Librosa · Docker · Federated Learning
Sageo
Collaboration · MOI LabsTrust and discovery layer for AI agents on MOI Chain.
A2A interaction proofs · on-chain agent identity registry
TypeScript · Node.js · MOI Chain · Google A2A Protocol · Coco contracts
Vision Transformer Paper Replication
Research · 2025From-scratch PyTorch implementation of ViT-Base from An Image is Worth 16x16 Words.
ViT-Base 86M params · ~93.8% test accuracy (pretrained ViT-B/16) · from-scratch vs pretrained comparison
Python · PyTorch · torchvision · torchinfo
Sprints & hackathons8 entries
FashionMNIST CNN Study
Progressive image classification study on FashionMNIST — linear baseline → non-linear layers → two-block Conv2d+MaxPool2d CNN. Reached ~88% test accuracy on 10-class classification, building CNN intuition before moving to pretrained architectures.
EfficientNet Feature Extraction & Experiment Tracking
Pretrained EfficientNet-B0 and B2 feature extractors on a 3-class food dataset — frozen backbone with a custom classifier head (3,843 trainable params from 4M+ frozen). Ran a systematic 8-experiment ablation across model scale, dataset size, and epoch count, tracked with TensorBoard and structured run-directory naming.
Agent Safe (ETH Denver 26)
Sandboxed execution environment for coding agents — scoped filesystem access, syscall filtering, and blast-radius containment for untrusted tool use.
GitHubCosmosPool (ETH Denver 25)
Cosmos SDK staking pool with delegator analytics and automated reward compounding via a governance-gated rebalancer.
GitHubEntreLink
Founder-to-founder matching platform — semantic search over founder profiles to surface co-founders, investors, and early hires.
GitHubEstateWise
Modular AI employee system for real estate transaction automation — MCP servers handling document requests, title review, and deal coordination.
GitHubFormulaOS
AI-native spreadsheet desktop app — formula engine, GPT-4 copilot, command palette, and Tauri desktop packaging.
GitHubClio (MBC 25)
Base L2 protocol for artist bonding-curve tokens — Solidity contracts with a social discovery layer for early fan-to-artist engagement.
GitHubExperience6 roles
SolstenSoftware 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
MOI Labs × BoilerBlockchainSoftware Engineer / Project Lead
Agent identity and interaction verification layer on MOI Chain.
Led SDK development for Sageo — built TypeScript wrappers for outgoing A2A clients and server-side request handlers that inject trace metadata and log hashed request/response proofs on MOI Coco contracts. Designed ephemeral-state fallbacks for devnet edge cases, wrote E2E workflow scripts to verify on-chain proof integrity, and co-authored the technical specification.
TypeScript · Node.js · MOI Chain · Google A2A Protocol · Coco
Kong LabsUndergraduate 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
BoilerBlockchain, PurdueHead of Education
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
ManoCo-founder / Software Engineer
AI-native deal coordination platform for commercial real estate acquisitions.
Co-founded and built Mano — an AI deal coordinator that automates the diligence process for CRE acquisitions. Built frontend flows in Next.js, backend API and data layer on Supabase, and the AI agent system handling document requests, estoppel collection, title review, and red-flag detection. The agent pre-empts every diligence task in parallel the moment an LOI is signed.
Next.js · TypeScript · Supabase · OpenAI · Resend · Tailwind
BoilerBite · Purdue SPARKCo-founder / Software Engineer
Built backend infrastructure for a student startup through Purdue's SPARK accelerator.
Co-founded BoilerBite through Purdue's SPARK entrepreneurship program. Designed and shipped the backend from scratch — Node.js/Express REST API, PostgreSQL with role-based access control, Clerk authentication, input validation with Zod, and production deployment on Purdue's Linux infrastructure via NGINX and PM2. Replaced an initial Firebase prototype with a robust, auditable API layer.
Node.js · Express · PostgreSQL · Clerk · Zod · NGINX · PM2
StackLanguages, tools, habits
PythonTypeScript
Node.jsFastAPIPostgresRedisDockerSupabaseRailway
PyTorchvLLM
DSPyHugging FaceBGEsentence-transformersVertex AIGemini
OpenTelemetrypytest
Git worktreescommand allowlistsCI/CDFoundryLinux
SolidityFoundry
EVMHedera HCSIPFSMOI ChainHardhat
ReactNext.js
TailwindTypeScriptFigmaWordPress
writing design docseval-firstsmall commitsread the sourceobservability before features
Open questions7 active threads
What should an agent remember?
Experimenting with TTL, episodic memory, and retrieval policies. Long-term memory becomes noisy surprisingly fast once workflows branch and tools start mutating shared state.
How much isolation do coding agents actually need?
Comparing worktrees, command allowlists, containers, and microVM approaches. Trying to understand where the security/ergonomics tradeoff stops being worth it.
Retrieval evaluation beyond MRR
Most retrieval benchmarks reward overlap, not usefulness. Interested in evaluation methods that survive paraphrase, ambiguity, and adversarial rewrite.
Compact execution traces
Exploring how much workflow history a verifier actually needs. Full traces become expensive quickly, but oversimplified summaries destroy reproducibility.
Shared state for autonomous systems
Looking at where distributed coordination actually helps agent systems: permissions, replayability, attribution, and conflict resolution across multiple executors.
Local-first agent workflows
Trying to understand which parts of autonomous workflows benefit from the cloud and which parts only ended up there because GPUs used to be inaccessible.
Trust boundaries in agent systems
Who is trusted to do what? Most agent frameworks blur the boundary between orchestration, execution, memory, and verification far more than they should.
Latest notes
SmartInvoiceExtractor
Regex extracts high-confidence fields from pdfplumber-parsed PDFs; Gemini fills gaps when critical fields are absent. A quality gate, JSON normalisation layer, and regex suggestion loop keep the LLM call rate low and the patterns improving. Production metrics are from the proprietary Solsten deployment — the public repo is an earlier local prototype.
Memora
Encrypted payload storage, consensus-ordered event log, and challenge-response permissioned key release — a tamper-evident memory architecture for agents that need to prove what they remembered and when.
Retrieval Adaptation Pipeline
A 3,292-example procedural disruption benchmark from YouCook2 and WikiHow, and a config-driven evaluation pipeline covering retrieval augmentation, reranking, and VLM strategies — with persisted retrieval and generation traces for reproducible multi-method comparison.
All notesDesk log
- neetcode-submissions: Add: lowest-common-ancestor-in-binary-search-tree - …
- neetcode-submissions: Add: subtree-of-a-binary-tree - submission-4
- neetcode-submissions: Add: same-binary-tree - submission-2
- AgentOS: Add SKILL.md skills executor with agentic tool loop …
- AgentOS: Add persistent agent memory with recall-at-task-star…
Updated 12 June 2026
Field notes
Most AI systems don't fail because the model is weak. They fail because nobody can see what happened between retrieval, prompting, tool execution, and state updates.
Memory is becoming more important than raw model size. A smaller model with stable retrieval and consistent state management is usually more useful than a larger one that forgets why it made a decision three steps ago.
Most multi-agent systems eventually rediscover distributed systems problems: retries, consensus, partial failure, stale state, and conflicting writes. The hard part is coordination, not generation.
Credentials & signalsVerified
- Purdue CSB.S. in Computer Science
- Machine Intelligence + SecurityCoursework focus
- Boiler Blockchain TATeaching Assistant
- Production SystemsShipped, monitored, iterated
Send the interesting thing.
I'm looking for Backend SWE, blockchain, and AI internships for Summer 2026 — particularly teams working on smart contracts, protocol infrastructure, retrieval, agent systems, or production AI platforms. If you're building something that needs to work past the demo, get in touch.
Fastest response via email or LinkedIn DM. Usually under 24h.