Applied AI & Agentic Systems
Agent traces connected to reviewed failures, regression cases, and release decisions.
Explore ControlSurface
Applied AI Engineer
Dev Chiniwala.
Building AI systems that think, remember & act.
At the intersection of agentic AI, memory systems, and software built for the real world.
Based in India
About me
Curiosity is the starting point.
I’ve always been curious about how things work, what makes them fail, and how to build them better.
That curiosity drew me to AI: agents that can reason, remember, adapt, and operate in the real world. I enjoy working across research, infrastructure, and product engineering to turn ambitious ideas into reliable systems.
I care about what happens after a convincing demo: whether a system works, scales, and lasts. My approach is a short feedback loop.
- Prototype
- Break
- Evaluate
- Harden
- Repeat
AI Memory & RAG Infrastructure
Persistent context that preserves provenance and accounts for facts changing over time.
Explore MNEMOSFull-Stack Product Engineering
Reliable APIs and considered interfaces built around the people using the software.
Explore Faculty recruitment portalExperience
Where I’ve been building.
From product engineering and network research to applied AI systems.
- May 2026 — PresentCurrent role
Applied AI Engineer
Stealth AI StartupCognitive AI & autonomous agents
Engineering cognitive AI systems, self-healing agents, and persistent memory architectures for production use.
- Python
- FastAPI
- LangGraph
- LangChain
- PostgreSQL
- Redis
- Docker
- RAG
- Next.js
- AI/ML
Work & responsibilities
- Architect agents with reasoning, memory, planning, and autonomous decision-making capabilities.
- Engineer workflows with failure detection, adaptive recovery, and iterative self-correction.
- Design persistent and temporal memory for context-aware, long-running agents.
- Build multi-agent orchestration and grounded retrieval into production AI systems.
- Specific product implementations and architecture details remain confidential.
- Jun — Jul 2026
LUSIP Hybrid Intern
LNMIITProduct engineering
Built a faculty recruitment portal connecting résumé parsing, Crossref and Scimago publication checks, and role-based application workflows.
- NLP
- REST APIs
- RBAC
- PDF generation
Work & responsibilities
- Developed résumé parsing and structured academic data entry for candidate applications.
- Integrated Crossref and Scimago data into a deterministic publication verification workflow.
- Implemented candidate and administrator permissions, application workflows, and PDF document rendering.
- May 2024 — Jun 2025
Full Stack Product Engineer
TRISX TechnologiesEdTech / FinTech
Built an online examination platform, REST APIs, and authenticated workflows across EdTech and FinTech products, with attention to SEO.
- JavaScript
- React
- Node.js
- Express
- MongoDB
- SEO
Work & responsibilities
- Built an online examination platform with responsive candidate workflows and dashboards.
- Developed REST APIs with Node.js, Express, and MongoDB, including JWT authentication and role-based access.
- Led design systems, wireframes, and responsive interfaces across EdTech and confidential FinTech work.
- Worked on SEO and search discoverability for product interfaces.
- May — Jul 2025
Researcher
LNMIITNetwork simulation
Modeled misinformation propagation with NetLogo, Barabási–Albert networks, and a multi-wave SIRS framework.
- NetLogo
- SIRS
- Network modeling
Work & responsibilities
- Developed a Barabási–Albert scale-free network model in NetLogo.
- Used a multi-wave SIRS framework to explore misinformation spread, fact-checking behavior, and reputation-sensitive social dynamics.
Community & leadership
Creative Lead & Machine Learning
GDG on Campus · LNMIIT
2024–2025Technical Lead
E-Cell · LNMIIT
2023–2024Selected projects
A few things I’ve built.
Independent tools, memory infrastructure, and products shaped by real engineering problems.
Agent production engineering
ControlSurface
Turn agent failures into reviewed regression tests.
An engineering control plane that turns agent traces into reviewed regressions and inspectable release evidence.
Design choiceAcknowledge only after a durable inbox write.
1,245.94
spans per second
Measured on one local WSL2 / Docker node with 10,000 synthetic spans; not a production capacity claim.
SourceOpen ControlSurface case study
Public · Release candidate
Temporal memory infrastructure
MNEMOS
Agents need memory that stays useful as facts change.
A self-editing memory runtime with bi-temporal validity, six retrieval signals, and token-aware context packing.
Design choiceSeparate temporal validity from retrieval relevance.
6
retrieval signals
Semantic, lexical, graph, time decay, cognitive, and memory tier; an architecture count, not an accuracy score.
SourceOpen MNEMOS case study
Public · Research runtime
AI browser workspace
PagePilot
Keep the webpage in the conversation.
A Chrome extension for webpage summarization and grounded conversation, backed by a RAG service.
Design choiceKeep model credentials behind the backend API.
More from the workbench
Products, tools, and ongoing explorations.
Verified intelligence systems
CortexOS
A local intelligence application pairing agent research with source-linked claims and a deterministic Rust verification path.
3
source-matching methods
SourceSee CortexOS in action.Load the recording to explore the desktop app.
Agentic product search
ShopSageAI
A conversational shopping agent using semantic routing, LangGraph workflows, and distinct product and policy retrieval paths.
University product engineering
LNMIIT Recruitment Portal
A faculty recruitment platform with résumé parsing, academic publication verification, role-based workflows, and document generation.
Persistent personal agents
HeyClara
A single-process personal assistant with scheduled jobs, persistent workspaces, memory consolidation, and messaging integrations.
24
documented MCP tools
SourceAdaptive learning intelligence
VYGOT
A private team project exploring personalized learning intelligence. Public implementation details remain limited ahead of launch.
Multi-agent exploration
NextGen-Trader
A Python and LangGraph code project exploring agent orchestration for trading analysis.
Content workflow automation
NexusFlow
A modular n8n workflow project connecting content generation and media services. Individual integrations require provider setup.
Open source
Small changes. Shared progress.
Fixing the details that other engineers depend on — retrieval performance, reliable SDKs, safer logs, and browser verification.
75
Read contribution notes Merged pull requeststo repositories outside my GitHub account
Reticle#1286
MergedMake browser verification respect declared endpoints.
Threaded named network URLs through action verification and replay so unrelated background requests do not distort the verdict.
Aligned verdict paths and added regression coverage.
TypeScript · Browser verificationChunkHound#419
MergedRetrieve missing embeddings without full-table scans.
Replaced whole-table chunk reads with targeted SQL and moved synchronous database work off the event loop, preserving the LanceDB fallback.
Reduced unnecessary reads; pinned behavior with contract tests.
Python · Retrieval performanceHelixDB#1164
MergedEnforce integer contracts across SDK boundaries.
Why this change matters
Added signed 64-bit range validation to Python and TypeScript SDK constructors, query values, and typed parameters.
Invalid values now fail locally; boundary cases have regression tests.
Python / TypeScript · SDK reliabilityDataMind#21
MergedKeep secrets out of audit error records.
Why this change matters
Improved audit error redaction to protect secrets while retaining useful debugging context.
A focused security fix at the logging boundary.
Python · SecurityGitHub snapshot · 16 Oct 2026. Count covers merged PRs authored by DevChiniwala, excluding repositories owned by that account. Each selected PR links to its merge record. View the GitHub search ↗
Engineering toolkit
From the model to the product.
The tools make more sense in context. Start with the systems I've built, or explore the full working inventory.
Durable AI infrastructure
ControlSurface- Python
- FastAPI
- PostgreSQL
- ClickHouse
Memory & retrieval
MNEMOS- Python
- Neo4j
- FAISS
Agents & orchestration
HeyClara- Claude Agent SDK
- MCP
Interfaces & product
PagePilot- React
- TypeScript
- FastAPI
Explore the complete toolkitSix areas · tools, practices, and project examples
A self-described working inventory across projects, experiments, and learning. Project examples show specific usage; inclusion does not imply equal proficiency with every tool.
The complete toolkit
Tool links open documentation
01Foundations
Languages & core
From system boundaries to product interfaces.
- Systems
- Services
- Product
In practiceControlSurface
ControlSurface connects a Python SDK and ingestion service to a Next.js investigation interface. Its PostgreSQL inbox and ClickHouse projections make the boundary between accepted and derived data explicit.
Read the ControlSurface case study02Intelligence
Models & AI
Training, inference, and the questions in between.
- Train
- Infer
- Evaluate
In practiceAdversarial ML reading notes
My adversarial-defense-methods repository curates research on attacks, defenses, robustness, and privacy. It is a reading collection; the notes do not imply a benchmark or a trained model.
Browse the research notes (opens in a new tab)03Orchestration
Agents & context
Connect reasoning to tools and useful evidence.
- RAG
- Tool calling
- Hybrid retrieval
- BM25
- Knowledge graphs
- Model evaluation
- Evidence-backed routing
- Guardrails
In practiceHeyClara
HeyClara connects the Claude Agent SDK and shared MCP tools in one persistent daemon. Scheduled work, messaging channels, and durable workspaces preserve continuity between conversations.
Read the HeyClara case study04Execution
Runtime & product
Take a capability from a process to an interface.
- Ray
- FastAPI
- vLLM
- SGLang
- Docker
- Kubernetes
- MLflow
- Jupyter
- React
- Next.js
- Tauri
- Node.js
- Express
- Tailwind CSS
- Motion
- Figma
- REST APIs
- Web
- Desktop systems
In practicePagePilot
PagePilot pairs a React browser extension with a FastAPI retrieval backend. Model credentials stay behind the API while the extension keeps the current webpage close to the conversation.
Read the PagePilot case study05State
Data & memory
Keep context useful as facts and time change.
- Temporal state
- Bounded context
- Traceable context
- Durable state
In practiceMNEMOS
MNEMOS separates temporal validity from retrieval relevance. Graph associations, dense and lexical retrieval, and token-aware packing help assemble context under a bounded budget.
Read the MNEMOS case study06Delivery
Infra & reliability
Make changes observable, reviewable, and repeatable.
- Regression tests
- Release gates
- Observability
- Incident intelligence
In practiceControlSurface
ControlSurface connects OpenTelemetry traces to incident review, regression cases, and release gates. Source evidence stays linked to the decisions made from it.
Read the ControlSurface case studySimulation & modeling
NetLogo documentation (opens in a new tab)Education
A foundation for what’s next.
The LNM Institute of Information Technology
Integrated B.Tech + M.TechComputer Science and Engineering
Five-year integrated dual degree, alongside applied AI research, open-source engineering, and campus technical leadership.
LNMIIT · Jaipur, India2022 — 2027
Selected certifications & continued learning
Complete Data Science, ML, Deep Learning & NLP Bootcamp
Udemy
Intro to SQL
Kaggle
Advanced SQL
Kaggle
Elements of AI
Elements of AI
Contact
Let’s build something meaningful.
Working on an ambitious AI product, agent infrastructure, or a difficult engineering problem? I’d love to hear about it.
Email me
View résuméOpen to interesting collaborations