1234567

y: systems built · x: systems understoodhuman → software → intelligent systems

Don't read my resume.
Talk to it.

Fig. 1 · Priyanshu Yogi, Jaipur, India. Notes.

Economics and political science, then self-taught software engineering, enterprise AI products, a lot of time with customers, and now agents, RAG and voice. Ask my AI about any of it.

Seven turning points, not seven job titles. Hover or click any point on the curve and the note updates. The ringed points are the first and the last: the dip between them is real, building fell while he learned the commercial side, and the last point is the correction.

Why am I relevant to your role?

Table 1Two educations, one method.

Formal training in how human systems behave; self-directed training in how to build technical ones. The first is why the second was learnable.

Human systems → software systems → intelligent systems
FormalSelf-directed
Programme
B.A. Humanities (Economics & Political Science)
University of Delhi, 2019 – 2022
Full Stack Web Development, Job Bootcamp
Coding Ninjas, 2022 – 2023
Ongoing
Self-directed, through projects and professional work, 2023 – present
What it trainedAnalytical thinking · Incentives · Institutions · Human systems · Why people and organizations behave the way they doJavaScript / TypeScript · Frontend · Backend · Databases · Cloud (AWS, Azure, Terraform) · System design · Computer vision solutioning · Edge deployments · LLMs · Agents · RAG · Voice systems
Carried forwardReading incentives before reading requirements. Treating an organization as a system with its own equilibrium.Learning a stack by shipping in it. Comfort with being the least credentialed person in the room and the most prepared.
  1. Economics + Political Science
  2. Curiosity about technology
  3. Learning to code
  4. Full-stack development
  5. Building real software
  6. Applied AI / Computer Vision
  7. Enterprise AI systems
  8. Agents with published evals
  1. aDates and programmes from the resume. No CS degree; nothing here is a certification list.
  2. bThe message is not “studied something unrelated and became an engineer.” It is “learned to think about human systems first, then taught himself to build technical ones.”

Fig. 2Experience that overlaps rather than stacks.

Six regions, one person. Click a region for its evidence, or ask the AI and watch the relevant ones hatch.

BuilderAgentsCustomerProductBusinessHumanities

Builder

Software engineering · Frontend · Backend · Integrations · System architecture · AI products

  1. 1Built frontend dashboards, backend APIs and Computer Vision pipeline integrations on HawkVision's core engineering team; used Terraform for deployment pipelines. Software Developer, HawkVision AI
  2. 2Built Instipe, a NestJS/GraphQL payment system for educational institutions processing 50+ lakh INR/month; cut a GraphQL query from 10+ seconds to under 100ms. SDE Intern, Edviron
  3. 3Led a 5-engineer team shipping Hazo (NestJS, React Native, AWS) and WorkOS. Entrepreneur in Residence, Project Dark Horse
  4. 4ReadTrail: a Manifest V3 Chrome extension with a canvas reading guide, a side panel and a writing Desk, every message validated at one service-worker trust boundary, 394 unit and behavioral tests and 11 Playwright end-to-end scenarios. ReadTrail
  5. 5SignalRoom: FastAPI and LangGraph with a SQLite checkpointer, strict-schema structured output recorded and replayed by content hash, and an evaluation suite that runs in CI without a key. SignalRoom
  6. 6Kept the portfolio's agent answering when a provider withdrew the free model it ran on, by routing across four OpenAI-compatible providers with a circuit breaker and a prompt sized to each one's token budget. Priyanshu OS / AI Voice Portfolio
  7. 7Spotify Mood Curator: Claude proposes songs, the listener's own Spotify data labels them, and a verify step checks what was saved against the plan. Two live CLI runs saved 25 of 25 and 30 of 30 songs in planned order with no duplicates. Spotify Mood Curator

A non-linear path, in six moves.

  1. 1

    Humanities

    Studied Economics and Political Science at the University of Delhi. Developed an interest in understanding systems, incentives, institutions and why people or organizations behave the way they do.†

  2. 2

    Self-taught engineering

    Moved into technology through self-learning: a full-stack bootcamp, then projects and professional work rather than a traditional CS degree.†

  3. 3

    Building

    Started working on software products: a payment system at Edviron, then leading a small team shipping Hazo and WorkOS at Project Dark Horse.††

  4. 4

    HawkVision

    Joined early on the small product/engineering team. Built parts of the product, worked on deployments, moved closer to customers, became Solutions Architect in June 2025 while already leading the solutions function and India sales, and in April 2026 was promoted to Head of Global Solutions and Head of India Sales.††

  5. 5

    Building in public

    Since leaving HawkVision in August 2026, building LLM agents and tools in the open, with the evals published, misses included.†††

  6. 6

    What comes next

    Between roles and available immediately, remote or on-site, for technically difficult products that solve real user problems.

  7. Roles he is looking at: Forward Deployed Engineer, AI Solutions Engineer, AI Engineer, Applied AI Engineer, Product Engineer, Customer Engineer, Solutions Architect, Technical Product roles.

Fig. 3Four roles, three companies, one direction.

Backend under real load, then 0→1 product leadership, then inside an enterprise computer-vision company from engineering to global solutions. HawkVision is the deepest chapter, not the only one. The last row is now: between roles, building in public.

Click any row to open what that role actually involved.

201920202021202220232024202520262027
Aug 2026 – present · Jaipur, India

Building in public, between roles

Independent

Left HawkVision in August 2026. Since then, building LLM agents and tools in the open with their evals published, misses included, while looking for the next role. Available immediately, remote or on-site.

  1. 1SignalRoom: a discovery-call transcript becomes a solution brief where every claim cites the line the customer said it on, with the evaluation and its misses published.
  2. 2ReadTrail: a privacy-first Chrome reading companion; 2.0 is on GitHub with 394 unit tests and 11 end-to-end scenarios, not yet on the Chrome Web Store.
  3. 3This portfolio: an agent grounded in six JSON files, with failover across four free-tier model providers and a prompt sized to each one's budget.
  4. 4Smaller experiments: Cochlea, a vocal-analysis MCP server heading toward a coach for beginner singers; Spotify Mood Curator, working and open source; and Atlas, a personal capture pipeline.
What it taught him

Evals are only worth publishing with the misses in them, and anything built on free tiers has to assume every provider will fail on its own schedule.

Technologies
PythonFastAPILangGraphTypeScriptNext.jsChrome Extensions (Manifest V3)MCPEvals
  1. aAll figures (camera counts, PoC counts, query timings) are as stated on the resume; none are rounded up here.

Table 2Not all projects are equal, so they aren't shown equally.

Featured products get full case studies. Technical projects are shipped systems from professional work. Experiments are labelled as what they are.

ReadTrail's ruler guide highlighting a line of the Wikipedia article 'Close reading' in pale red

ReadTrail

Working MVP, 2.0 on main and unreleased; not yet on the Chrome Web Store

A privacy-first Chrome reading companion: follow the line, keep your exact place, and write beside the passages you saved.

Explore ReadTrail GitHub
0
The portfolio homepage: a career plotted as Fig. 1 on quadrille paper, with an ask input drawn as an annotation on the plot

Priyanshu OS / AI Voice Portfolio

Working MVP: deployed on its own domain, and you are reading it

A portfolio you talk to. An AI agent grounded in structured profile data that can answer, navigate and reason about fit.

Explore Priyanshu Live demo GitHub
1
The SignalRoom review workspace: changed claims, the brief on a line-number gutter, and the review panel

SignalRoom

Working demo with a published evaluation, synthetic data only

A discovery transcript becomes a solution brief where every claim is tied to a line the customer said, a critic lists what could not be supported, and a human approves.

Explore SignalRoom Live demo GitHub
2
Technical projectWhat it didWhereStatus
AI-native solutioning operating system
Claude · Obsidian · Research automation · Knowledge management
Claude + Obsidian system that automated company, process and operations research for enterprise CV proposals.Proposal first drafts in 5-10 minutes; Solution Architecture Spec first drafts in under 5 minutes; customer and market intelligence auto-routed into a shared knowledge vault.HawkVision AIJun 2025 – Aug 2026 · promoted Apr 2026Shipped internally at HawkVision
WorkOS
Google Calendar API · Jira integration · Role-based access · Agentic workflows
An agentic enterprise operating system unifying role-based access, onboarding and cross-tool integrations.One of two products shipped by a 5-engineer team from an ambiguous early-stage vision.Project Dark HorseJun 2024 – Mar 2025Shipped at Project Dark Horse
Hazo
NestJS · React Native · AWS
Mobile product with a NestJS backend, React Native frontend and AWS infrastructure.Led the team building it; directed roadmap and delivery cadence for 30% faster time-to-market.Project Dark HorseJun 2024 – Mar 2025Shipped at Project Dark Horse
Instipe
NestJS · GraphQL · MongoDB · AWS Lambda
NestJS/GraphQL payment system for educational institutions processing 50+ lakh INR/month.MongoDB pipeline optimisation cut query execution time up to 90%; one GraphQL query went from 10+ seconds to under 100ms with caching and AWS Lambda.EdvironNov 2023 – May 2024Shipped at Edviron
Highway-safety Computer Vision PoC
Computer Vision · Camera placement strategy · Edge deployment · Solution Architecture Spec
10-camera highway-safety PoC across a 165km corridor on on-prem edge hardware, no cloud, with a path to ~100 cameras.Owned discovery, feasibility, camera placement, architecture and deployment planning.HawkVision AIJun 2025 – Aug 2026 · promoted Apr 2026PoC delivered, path to ~100 cameras
  1. e1Spotify Mood Curator [working mvp, open source] A Claude skill and Python CLI that builds mood playlists from real Spotify listening: Claude proposes the songs, the listener's own data labels each one played, saved, known or new, and a verify step checks what was saved against the plan.
  2. e2Cochlea [in progress, repo private for now] A local MCP server that measures a sung take (pitch in cents, drift, vibrato, phrases, key, tempo) so an LLM coaches from cited, time-stamped numbers. Next, not built yet: a coach for beginners who like to sing but know no music theory.
  3. e3Atlas [in progress, architecture only] A personal capture-and-triage pipeline: Telegram into a Cloudflare Worker running a D1 job queue with leases and retries, daily per-provider model budgets reserved atomically, a privacy gate that keeps personal content on no-training providers, and approve-by-tap before anything reaches Notion.
  4. e4Agentic RAG experiments [exploring] LangGraph, vector databases, memory, tools and structured workflows.
  5. e5EdgeVision Lite [planned] A real-time computer-vision system: video streams, detection, tracking, rules, events, APIs and a local dashboard. Bringing the CV solutioning experience back into hands-on code.
  6. e6ANPR video pipeline [planned] Video processing with timestamped detections and structured exports.

Fig. 4A background that connects, not a list of keywords.

Drag nodes. Click one to see what it links to. Economics sits three hops from a highway camera deployment for a reason.

Experience

HawkVision AI

Open Solutions Architect → Global Solutions Head & Head of India Sales, HawkVision AI
Connected to

Table 3Why am I relevant to your role?

Paste a job description. The agent reads it against Priyanshu's structured experience and returns strong matches, transferable experience and the gaps. Every match cites its source row. A fit report with no gaps is not trustworthy, so gaps are the point.

0 characters · not stored
The ledger has five rows: strong matches, transferable experience, gaps, relevant projects, and questions worth asking him. Nothing is stored.

Fig. 5Things I keep going down rabbit holes on.

Not a skills list. Solid points are things he works on; hollow points are things he is curious about. Positions are a self-portrait, not data. Drag them if you disagree.

Hover or tab to any point to read the note behind it.

builds withreads about