Open to Protocol Risk Analyst & DeFi risk roles

Abdul Rahman

Protocol Risk Analyst  ·  Perps & DeFi risk  ·  Monitoring & Reconciliation

I work on how a DeFi protocol can lose money, and on proving it hasn't: margin and liquidation rules, funding, oracle and vault safeguards, live solvency checks against the chain, and the monitors and alerts that watch them around the clock. I write the spec, build with AI tools, and verify the result before it ships.

The main projects are live and linked, so you can open them rather than take my word for it.

About

A bit about me.

Over the last two years I've worked on a live Solana perps exchange, on the side of it that has to be right: margin and liquidation mechanics, oracle and vault safeguards, reconciling what the exchange reports against the chain, and the monitors and alerts the team relies on. I've also built agent and retrieval systems with proper guardrails.

The common thread is that these systems are trusted to act, so they have to be right. On a trading platform a wrong output isn't an awkward answer, it's someone's money.

I build by directing AI coding tools. I write the spec, set the acceptance criteria and the tests, and reject whatever fails them. Model output is treated as untrusted until it passes those checks.

Domains worked in4
Years on production AI systems2+
Public repositories26
Apps linked on this page6

Experience

Where I've worked.

Jan 2025 — Aug 2026

Protocol Risk Analyst

Flash.trade

  • ▸Specified a vault monitor that ran around the clock and checked every deposit and withdrawal against the chain, using two independent data sources so a single bad feed can't slip through.
  • ▸The monitor I'd set up flagged a public $98K over-withdrawal exploit as it happened and traced the attacker's full withdrawal history from chain data. The protocol team handled the fix and users were reimbursed.
  • ▸Set up an independent price check that compared the exchange's mark price with an external Pyth feed and alerted when they drifted more than 1.5% apart.
  • ▸Checked the exchange's reported figures against the chain: decoded all 898 trading accounts and tested the 165 open positions against the protocol's own rules. All of them passed.
  • ▸Tracked the V1-to-V2 migration from chain data. 894 of 898 accounts had moved and the remaining four held no positions, so the cutover could be signed off.
  • ▸Built the dashboards the team works from day to day: pool utilisation, custody balances, and position and collateral reconciliation, all read from chain state.
  • ▸Built a chat interface for trading across 34 markets: limit orders with take-profit and stop-loss, trigger orders, staking, yield and transfers. I looked at how people actually phrase requests and wrote the routing rules from that, so common commands don't need a model call.
  • ▸Built a command-line terminal that takes plain-English instructions instead of clicking through the website. The same instruction runs the same way every time, with an automated test suite behind it.
Jul 2024 — Dec 2024

Business Analyst Intern

Flash.trade

  • ▸Wired live price and position feeds into the trading terminal, with reconnect handling and stale-data checks so it never showed a number it couldn't back up.
  • ▸Built market-data dashboard components: market lists, position tables and pool views, all reading on-chain state.
  • ▸Wrote automated tests for the data layer. The later production suite grew out of that.
  • ▸Documented the on-chain data model (account layouts, decoding rules, what each field means), which the later dashboards were built from.
Jun — Jul 2023

Data Science Intern

SkillDzire — Hyderabad, India

  • ▸Completed a structured data-science training programme covering Python, statistical analysis and applied machine learning on real-world datasets.
  • ▸Worked through supervised-learning exercises in linear regression and decision trees, and their evaluation.
  • ▸Applied Pandas, Matplotlib and Seaborn to guided data-cleaning, exploratory-analysis and visualisation tasks.

Projects

Things I've built.

They share one rule: a model can propose, but something deterministic checks it before anything acts.

Archived · Jul–Sep 202601

Flow Sentinel — autonomous monitor

A monitor that ran unattended on Flash Trade's vaults until the protocol wound down in September 2026: hourly outflow limits, balance checks to the exact unit, and an independent oracle cross-check, with an alert when its two data sources disagreed. Every figure traced back to a real transaction. The archive keeps the final guard thresholds and the record of what it caught.

What it shows: monitoring that ran without supervision, and an honest record

Real-timeDual-witnessOracle cross-checkAlerting
Open the archive →
Archived · Jul 202602

Flash V2 On-Chain Census — reconciliation

Decoded every basket, position, order, custody and market account of a live perps exchange from its rollup with the program's own on-chain IDL, and reconciled them to the protocol's totals to the raw integer: 64 of 64 market sides exact, 508 positions in 433 baskets across 10 pools, and 27 of 27 vaults backed with zero deficit. It refreshed every 10 seconds while the exchange was live; the archive keeps the final verified scan from 7 July 2026.

What it shows: an exchange's reported state checked against the chain

SolanaReconciliationVault solvencyOn-chain IDL
Open the archive →
npm03

Builder Terminal

Turns plain-English commands into live trades across 34 markets. Model output goes through a deterministic parser and per-trade signing guards before anything is sent, so a model mistake can't move funds on its own. Audit log, program allowlist, kill switch.

What it shows: AI actions with hard checks in front of them

TypeScriptAgent loopsSigning guardsAudit log
View source →
Live04

Polymath — agentic RAG

Load a set of documents (guidelines, contracts, filings, a handbook) and ask questions against them, with cited answers. Switching domains needs no code change. The agent decides how many retrieval hops to take, searches by vectors and keywords together, and a second model scores each answer for faithfulness to the sources it cited.

What it shows: retrieval with its own evaluation built in

Next.js 16ClaudepgvectorLocal embeddings
Open live demo →
32 tests05

Specto — visual inspection

Defect detection you can talk to. A vision-language model finds defects with no training data, so it works from day one on images nobody has labelled. Your corrections train a private PatchCore / DINOv2 detector that gradually takes over.

What it shows: vision models that improve from corrections

VLMPatchCoreDINOv2Few-shot
View source →
All work→

20+ public repositories

Trading terminals, dashboards, monitors, AI tooling. Two packages on npm, one with a Homebrew tap. The rest is on GitHub.

Browse GitHub →

Case study · July 2026

The day the monitor had to work.

In July 2026 a publicly reported over-withdrawal exploit took $98,000 from the protocol. Users were reimbursed in full.

The monitor I'd built was already running, and it flagged the transaction.

To be clear about my part: I built the monitor, it detected the transaction and sent the alert. I didn't prevent the exploit or fix it; the protocol team did that.

DetectedThe monitor flagged the over-withdrawal from live chain data as it happened, not in a post-mortem.
ReportedA Telegram alert went out automatically with the transaction, so the people who could act on it knew straight away.
ConfirmedThe balance checks kept running throughout, so the team could see the actual vault balance while it was still unfolding instead of estimating the damage.
Why it existedI'd built it a few days earlier, after an $18M oracle exploit at another protocol showed how this kind of failure plays out.

Skills

What I work with.

Models

01

The models I work with day to day.

ClaudeClaude CodeGPT-5OpenAI CodexGeminiGrokLlamaLocal embeddingsModel selectionCost & latency tuning

Agentic & retrieval

02

Agents, retrieval, and keeping answers tied to their sources.

Agentic AIAgent loopsTool callingMCPRAGHybrid searchReciprocal rank fusionpgvectorEmbeddingsRerankingCitation groundingFaithfulness scoringPrompt engineeringAI guardrailsEvaluation

Multimodal

03

Vision models and image-based detection.

Vision-language modelsZero-shot detectionFew-shot learningPatchCoreDINOv2Anomaly detectionCorrection loops

Protocol risk & monitoring

04

Perps risk, monitoring, alerts and automation.

Margin & liquidation rulesFunding & borrow-rate modelsOracle safeguardsVault solvency invariantsExecution-quality measurementOn-chain reconciliationAI-agent workflowsAutomated reportingWorkflow automationAPI integrationWebhooksReal-time pipelinesScheduled monitoringTelegram Bot APIAlertingIncident response

Delivery & engineering

05

Scoping, delivery and documentation.

Requirements & scoping (BRD, FRS)SDLCUAT supportKPI & metric definitionData reconciliationRisk & exposure analysisStakeholder reportingProof of conceptClient-facing demosAgile / ScrumJIRAPythonREST APIsSQLPostgreSQLpgvectorGitVercel

Building fluency · 2026

06

Tools I'm currently learning. Ask me how far I've got.

n8nMake.comZapierOdoo ERPAzure AI FoundryCopilot StudioPower PlatformPower BILangChainLlamaIndexAzure OpenAI

Education

Where I studied.

2020 — 2024

B.E. Computer Science

Osmania University, Hyderabad

Four-year computer science degree, graduated 2024.

7.96GPA / 10
2024Graduated
4Years

Certifications

Certifications and training.

Completed credentials, plus three in progress for 2026. Mostly Azure, since that's what most enterprise AI work in the Gulf runs on.

CiscoTraining

CCNA — Networking Programme

Cisco Networking Academy

Routing, switching, firewall · 2024

Palo Alto NetworksEarned

Cloud Security Fundamentals

PANW

Cloud security principles · 2025

MicrosoftIn progress

Azure AI Apps & Agents Developer Associate

AI-103

Agentic AI on Azure · 2026

MicrosoftIn progress

Azure AI Fundamentals

AI-901

Azure AI services · 2026

Amazon Web ServicesIn progress

AWS Certified AI Practitioner

AIF-C01

AI/ML on AWS · 2026

ForageCompleted

Virtual Experience Programmes

AWS · Accenture · Deloitte

Solutions architecture, data analytics & visualisation, technology · 2023–2025

★
Recognition · 2018

Guinness World Record participant

Part of the Largest picture mosaic formed by people, an officially confirmed record set at International Indian School Jeddah.

4,500participants
Resume · PDF · 1 page · Updated Oct 2026

Abdul Rahman — Protocol Risk Analyst

The one-page version of this site. The links inside are clickable.

Contact

Get in touch.

If you run a protocol and want to know how it can lose money, have reports nobody quite trusts, or a system that needs watching around the clock, that's the kind of work I do. Have a look at one of the live links first, then drop me a note.