# Vamshi Jandhyala

> Making enterprise data and APIs agent-ready: the reliable, governable foundation that AI agents act on, after two decades in financial-services data and platforms.

Vamshi Jandhyala is a London-based product and technology leader with 20+ years at Bloomberg, JP Morgan, Fidelity, Schroders, HSBC, BlackRock, and PwC. His work sits at the intersection of AI product management, API design, and enterprise data platforms. He advises early-stage startups in the same space, and publishes mathematical books and handouts under his personal imprint.

## Work

- [About](https://vamshij.com/about): bio, career, education, patents, contact.

## AI Lab

- [Rule Grounding](https://vamshij.com/lab/rule-grounding): A working prototype of citation grounding for regulated agent output, built against the FCA Handbook API on the day after it launched. It resolves a citation, checks whether the provision is binding, and refuses to certify a decision whose governing text has changed since it was taken.
- [The reasoning was defensible. The answer was wrong.](https://vamshij.com/lab/arranged-data-eval): On a regulated-finance task, the same model produces two different answers because only one dataset encodes the firm's record of truth. A deterministic eval tied to that truth catches the difference, and no external benchmark can substitute for it.
- [ContextScope](https://vamshij.com/lab/contextscope): A clickable prototype for a causal debugger of agent state: what the agent believed, why it believed it, what conflicted, and what that enabled next. Where AgentScope makes a run legible, ContextScope makes the agent's belief legible. Built around the conflict a regulated operation has to handle well: the ledger says the card payment was authorised, the customer says it was not them.
- [Tax Policy Navigator](https://vamshij.com/lab/tax-policy-navigator): A case study in product judgment for grounded AI in a regulated domain. Why refusal has to be a designed surface, why a citation has to reach down to the individual claim, and where a machine judge stops being enough and a human has to take over. Built and evaluated end to end over the full HMRC Employment Income Manual. The hard problems turned out to be product problems, not model problems.
- [AgentScope](https://vamshij.com/lab/agent-harness): A clickable prototype that visualises a tree of agent runs, each with their own steps, context window, and cost share, on one screen. Three switchable mock runs demonstrate the design across different shapes.
- [Data Discovery Agent](https://vamshij.com/lab/datadiscoveragent): A clickable prototype that answers a question most chat-with-data demos refuse to ask. How should a regulated-data platform expose an agent surface when picking the wrong dataset can mean wrong answers, compliance breaches, or runaway cost?

## Writing

- [The authority surface](https://vamshij.com/writing/the-authority-surface): On your laptop you answer the question may-the-agent-do-this by being there to click yes. In production no one is, so what an agent may do alone, what needs a person, and what it must never do has to be decided in advance, action by action. Graded authority is a product decision, not a setting.
- [The record nobody agreed to keep](https://vamshij.com/writing/the-record-nobody-agreed-to-keep): Four days apart in July 2026, the EU postponed its record-keeping obligation for standalone high-risk AI by sixteen months and a widely adopted tool protocol deprecated its logging feature. Firms now have until December 2027 to produce a record nobody has agreed to keep. Each layer defers accountability to the layer next door, every deferral is locally correct, and the chain closes on nobody.
- [Encoding judgement](https://vamshij.com/writing/encoding-judgement): Agents rarely lack judgement because the model is weak. They lack it because the judgement was never written down anywhere the agent can reach. From the expert systems that tried to state judgement as rules to the preference data and constitutions that now capture it by example, and why the hard part has moved from articulating judgement to seeing what we encoded.
- [A benchmark is not a control](https://vamshij.com/writing/a-benchmark-is-not-a-control): The most respected LLM benchmarks measure whether a model can do a task in a lab. None of them measure whether an agent's output is acceptable in a regulated workflow, and most cannot, by construction. A field guide to reading a leaderboard when you run agents on data you have to answer for.
- [An answer that runs is not an answer you can trust](https://vamshij.com/writing/an-answer-that-runs): A generated query that runs is not one you can trust, and the gap widens as the data outgrows what a human can check by eye. What an agent harness for interactive data discovery has to do, told through a quant at a hedge fund: keep the analyst able to trust a result they can no longer inspect.
- [The cushions an agent loses in production](https://vamshij.com/writing/the-cushions): An agent that works beautifully on your laptop is standing on cushions you stopped noticing: a human watching, a single user, your own credentials, a visible trace, your own judgment, and an undo key. Production removes them one by one. Engineering them back is what reliability actually means, and most of it is not a modeling problem.
- [Vibe Coding and the Boundary of the Firm](https://vamshij.com/writing/coase-vibe-coding): Vibe coding has not only moved the make-or-buy boundary; it has turned building into a cheap option that reprices the whole decision, even for the core systems the cloud era taught everyone to rent. What should decide it is no longer the falling cost of code but the costs that do not fall with it, amortisation, accountability, data, and optionality.
- [Agent-ready financial data: an architectural view](https://vamshij.com/writing/agent-ready-financial-data): An architectural view of how a financial data provider can make its data ready for autonomous and semi-autonomous agents. Workflow archetypes, the three human-in-the-loop bands, the surfaces (REST, SDK, MCP, agents.md, skills, code-as-action), and why the semantic layer is the product.
- [Ironies of AI automation](https://vamshij.com/writing/ironies-ai-automation): Lisanne Bainbridge showed in 1983 that automating a task makes the human who remains more important and less capable at the same time. Every one of her ironies survives the move to AI agents, and two of them get worse.
- [The catalog becomes the query interface](https://vamshij.com/writing/catalog-as-query-interface): An essay from 2023 arguing that natural-language queries would replace SQL as the primary interface to enterprise data, with the data catalog as the substrate that makes those queries trustworthy. Republished with a 2026 note.
- [When the customer is an LLM](https://vamshij.com/writing/llm-apis-the-new-customer): An essay from 2023 arguing that LLMs would become primary consumers of APIs, that API documentation would become in-demand training data, and that API product managers would soon be designing for a non-human customer. Republished with a 2026 note.

## Mathematics

- [A Faulty Traffic Signal](https://vamshij.com/mathematics/faulty_traffic_signal): A 15-second pedestrian signal whose red/green split is itself uniformly random. Closed forms for the probability of red, the expected wait, and the expected total time to cross, each from a one-line integral, with a sample-space picture, an account of length-biasing, and a Monte Carlo check of every result.
- [Charlotte's New Web](https://vamshij.com/mathematics/charlottes_web): A circular billiard puzzle by Xavier Durawa: Charlotte's strands reflect inside a circular frame; how many earlier strands does the n-th strand cross on average? A self-contained walk through the geometry, the modular-arithmetic reformulation, the piecewise integration, the antipode trick, and the telescoping sum that lands the exact closed form.
- [Beautiful problems deserve beautiful typesetting](https://vamshij.com/mathematics/beautiful-problems): On the hunt for beautiful olympiad problems, the indignity of badly typeset mathematics on the web, and a small side project to surface AI training datasets back to the human readers the problems were originally written for.
- [Cozy Circles in Regular Polygons](https://vamshij.com/mathematics/cozy_circles): A puzzle by Xavier Durawa: in a regular polygon, place a circle at the midpoint of each side, tangent to the side and as large as possible without overlap. What fraction of the polygon's area is covered by the n circles?
- [The Apollonian Gasket](https://vamshij.com/mathematics/apollonian_gasket): Constructing the Apollonian gasket fractal from Descartes' Circle Theorem and its complex extension. Includes an elementary algebraic proof, the Lagarias-Mallows-Wilks complex form, and a queue-based Python implementation.
- [Appeasing the Cherry Blossom Horde](https://vamshij.com/mathematics/cherry_blossom_puzzle): A geometric-probability puzzle by Xavier Durawa: a random chord across a circle intersects a diameter; given that intersection, what is the expected ratio of the shorter segment of the diameter to the longer?
- [Posidoku](https://vamshij.com/mathematics/posidoku): A Sudoku variant by Alf Smith with no number clues, only positional gold cells whose values must equal their row, column, or box position. Solved with Google's CP-SAT.
- [Solving the Jumping Julia Maze](https://vamshij.com/mathematics/jumping_julia): A puzzle from the Julia Robinson Mathematics Festival: navigate from the top-left corner of a grid to the goal cell at the bottom-right, where each cell's number specifies the exact distance you may jump horizontally or vertically to the next cell.
- [Building a LinkedIn Tango Solver with Z3](https://vamshij.com/mathematics/linkedin_tango): A Z3 solver for LinkedIn's daily Tango puzzle: fill an n×n grid with suns and moons such that no three adjacent cells in any row or column share a symbol; row and column counts are balanced; and pairs of cells linked by = or × constraints match or oppose.
- [Maximising the Length of a Projectile Trajectory](https://vamshij.com/mathematics/projectile_trajectory): At what angle should a projectile be launched, under uniform gravity and no air resistance, so that the arc length of its trajectory is maximised? A standard integral and one implicit equation give the answer.
- [Subsets of {1,…,n} with Exactly One Pair of Consecutive Integers](https://vamshij.com/mathematics/subsets_consecutive_pair): A generating-function argument in which the count is the convolution of Fibonacci numbers; partial fractions over the golden-ratio roots give a closed form involving Fibonacci and Lucas numbers, with f(10) = 235.
- [Running Total of a Die](https://vamshij.com/mathematics/die_puzzle): Roll a fair six-sided die until the running total first exceeds 12. What is the most likely final total?
- [Solving the LinkedIn Queens Puzzle with Z3](https://vamshij.com/mathematics/linkedin_queens): LinkedIn's Queens puzzle: place one queen per row, column, and colour region of an n×n board, with the additional constraint that no two queens touch, not even diagonally.
- [Expected Distance of a Random Point from the Centre of a Regular Polygon](https://vamshij.com/mathematics/geometric_expectation_polygon): Closed form for the expected distance from the centre of a regular n-gon of unit circumradius, via a Jacobian substitution on one fundamental triangle; three triangle-sampling methods for computational verification.
- [Knights on a Chessboard](https://vamshij.com/mathematics/knights-on-a-chessboard): A white knight and a black knight on diagonally opposite corners of a 3×3 square. What is the expected number of moves until the black knight captures the white one? A clean Markov-chain problem with a closed form.
- [Inversion in Geometry](https://vamshij.com/mathematics/inversion): Inversion in a circle transforms two hard problems about tangent circles into routine ones: a Pappus chain (prove that the height of the n-th circle equals 2n times its radius) and the distance between the circumscribed and inscribed circles of three mutually tangent circles of radii 1, 2, 3.
- [Islamic Geometric Patterns](https://vamshij.com/mathematics/islamic_geometric_patterns): A computational construction of star patterns from the Islamic tradition: translational units, motifs, rosettes, and the rosette dual, implemented in Python with NumPy.

## Books

- [The Fiddler: Solutions](https://vamshij.com/books/fiddler): Independent solutions, verified in Python, to puzzles from Zach Wissner-Gross's Fiddler on the Proof, the successor to FiveThirtyEight's Riddler. Each chapter states the puzzle with attribution, solves it from scratch, takes on the extra credit, and closes with a short program that confirms the answer numerically.
- [Jane Street: Solutions](https://vamshij.com/books/janestreet): Independent solutions, verified in code, to the monthly puzzles set by Jane Street. Each chapter states the puzzle with attribution, solves it from scratch, and closes with the program that confirms the answer. Grid-filling puzzles go to a constraint solver; the probability puzzles are derived by hand and checked by simulation.
- [Monthly Mindbenders: Solutions](https://vamshij.com/books/momath): Independent solutions, verified in code, to Peter Winkler's Monthly Mindbenders for the National Museum of Mathematics. Each chapter states the problem with attribution, solves it from scratch, and closes with the program that confirms the answer. The problems reward one clean idea rather than machinery, which is what makes them worth writing up.
- [Nikoli](https://vamshij.com/books/nikoli): Twenty-five Japanese pencil puzzles, Kakuro, Sudoku, Skyscrapers, Kakurasu, Takuzu, KenKen, Flow Free, Slitherlink, Hashi, Akari, Shikaku, Nurikabe, Masyu, Hitori, Fillomino, Heyawake, Shakashaka, Marupeke, Walls, L-Panel, BlockNumber, Searchlights, Numbrix, and Three-in-a-Row, each modelled as a constraint-satisfaction problem and solved by Google's CP-SAT.
- [Number Puzzles](https://vamshij.com/books/number_puzzles): Fifty classical number puzzles, re-authored in contemporary language and given fresh, fully worked solutions, the suppressed digit and casting out nines, the taxicab number 1729, square palindromes and figurate numbers, anomalous cancellation, automorphic and self-describing numbers, the coconut Diophantine, Ramanujan's house number, the only magic hexagon, the eight queens, and more.
- [Solving Puzzles through Mathematical Programming](https://vamshij.com/books/puzzles_mp): Fifty-nine classical and contemporary puzzles, among them Fish, Calendar, Praxis Rhombus, Bedlam Cube, Instant Insanity, Drive Ya Nuts, Ostomachion, Langford, Quintomino on the dodecahedron, Dobble, Monkey Cat Dog, Prime Circle, Rolling Cubes, and Pilgrims, modelled as constraint-satisfaction problems and solved by Z3, CP-SAT, and NetworkX.
- [The Riddler](https://vamshij.com/books/riddler): Selected problems from Oliver Roeder's *Riddler* column at FiveThirtyEight, set out cleanly and solved with a mix of paper-and-pencil reasoning, generating-function techniques, and small Python simulations where the closed form is out of reach.
- [Problem of the Week: Solutions](https://vamshij.com/books/wagon): Independent solutions, verified in code, to problems from Stan Wagon's Problem of the Week. Each chapter states the problem with attribution, solves it from scratch, and closes with the program that confirms the answer, including the cases where the published problem does not say quite what it means to say.

## Optional

- [RSS feed](https://vamshij.com/rss.xml)
- [Sitemap](https://vamshij.com/sitemap-index.xml)