Enterprise Architecture
HLD/LLD, integration design, modernization, system boundaries and build-vs-buy decisions.
Forward Deployed Engineer / Applied AI Engineer
I turn complex business workflows into working software, from customer discovery and architecture to full-stack delivery and production integrations.
My work spans enterprise telehealth, live commerce and investment-research products, authenticated MCP tools, document retrieval and evaluated AI workflows. I use coding agents throughout development and own the engineering decisions, validation and release.

End-to-end engineering
Business thinking. System design. Hands-on delivery.
I connect the pieces — and own the outcome.
01 / DISCOVER
Turn a business idea into a clear scope. Understand users, constraints and what success should look like.
Additional expertise
Hands-on experience across product, enterprise systems, security boundaries, integrations and business domains.
HLD/LLD, integration design, modernization, system boundaries and build-vs-buy decisions.
Hybrid retrieval, authenticated MCP tools, evaluated AI workflows and development with coding agents.
OWASP, vulnerability research, Hack The Box, HackerOne, Kali Linux, identity and API security.
Trading and investment systems, financial data workflows, analytics and payment-related architecture.
Smart contracts, blockchain infrastructure, nodes, wallets, on-chain integrations and automation.
AWS, Docker, Kubernetes, CI/CD and production platform architecture.
SQL, NoSQL, search/vector systems, ingestion pipelines and data-oriented architecture.
Discovery, architecture, hands-on implementation, launch and iterative product delivery.
Independent products, enterprise systems and client projects. A closer look at the problems, decisions and delivery.
6 featured projects from a portfolio of 17.
Selected client engagements are anonymized to respect confidentiality.
No cases in this category.
Commerce, operations & AI
Independent Product
Retail, a B2B partner platform and an SMM agent, connected with business data, knowledge retrieval and MCP tools.
Retail, a B2B partner platform and an SMM agent, connected with business data, knowledge retrieval and MCP tools.
The challenge
Catalog, orders, inventory, financial records and advertising data sit across different services. Operators need consistent product data, traceable reports and useful AI tools that respect the same business rules as the application.
Originally built as Taxventory, an inventory management system, the platform evolved into Pets pets pets’ online store and operational backbone.
Product & engineering
A commerce ecosystem spanning a retail store, a B2B partner platform, business operations and an SMM agent. I own product decisions, architecture, full-stack implementation and delivery across these systems.
The original inventory application evolved into the storefront and operator platform. Catalog, orders, payments, delivery and channel synchronization share one product rather than a separate store rewrite.
Built a separate wholesale portal for verified companies: private catalog, bulk CSV ordering, carton/minimum quantities, partner tiers, individual discounts and payment terms.
Checkout uses authoritative server quotes, atomic reservations, idempotency and immutable order snapshots. Admin tools cover procurement, expiry-aware stock and FEFO shipment, buyback and consignment records, and configurable GAMA incentives. The platform has its own React interface, company-scoped access and PostgreSQL database, with commercial rules shared consistently between buyer quotes and administrator controls.
Built an editorial workspace with content pillars and weekly planning in PL/EN/DE/UK, defaulting to two slots a day.
Queue-backed AI generation prepares captions, scripts and image/video production briefs for Instagram, Facebook and TikTok. Editors request changes and feed comments into regeneration; immutable versions and server-computed approval hashes bind each decision to exact saved content. Edits invalidate approval, and nutrition/health content requires supplied sources and a recorded expert-review confirmation. Daily generation limits, stale-result rejection and cancellation keep background work under operator control.
Reports combine BaseLinker sales, Taxxo records and Allegro billing and advertising data.
Account scopes, net/gross basis, source dates and incomplete-data indicators remain explicit. Missing values stay unknown, and advertising-attributed sales are kept separate from company revenue.
Page-aware PDF ingestion preserves document and page provenance.
Hybrid search combines pgvector and lexical rankings through reciprocal rank fusion. OAuth with PKCE protects ten read-only MCP tools: two for document search and sources, and eight for business queries. Business responses reuse the operator UI’s reporting logic.
AI-assisted extraction, translation and authoring prepare catalog data for review.
Transfer tooling records a change preview, checks source fingerprints, verifies writes by reading them back, and supports journaling and guarded rollback. Model output is input to validation and review, not authority to publish.
System design
Discovery, checkout, payments and delivery.
Company terms, stock and wholesale ordering.
Finance, inventory and advertising sources.
Retrieval, MCP tools and editorial approval.
The retail Spring Boot / Angular application owns commerce, reporting and the SMM workflow. A Python/FastAPI service handles retrieval and MCP, delegating business queries to existing reporting services. B2B is a separate Spring Boot / React application with its own PostgreSQL database, company-scoped access and deployment. Retail and wholesale share product context without assuming shared prices, stock or customer sessions.
Outcome & evidence
Delivered outcome
Retail is available at pets-pets-pets.com and the B2B partner platform at b2b.pets-pets-pets.com. Pets-Pets-Pets Studio provides AI-assisted editorial planning, generation and human approval. Hybrid knowledge search and business-source status queries were verified through the connected MCP service on 14 November 2026.
B2B maintains its own commercial rules, stock and company permissions. The SMM workflow covers editorial planning, text, scripts and production briefs with human approval. Internal financial and customer data are omitted from this public case.
Evidence to approved Task
AI Lab Project
Structured LLM extraction, rule retrieval and a local FHIR sandbox, with human-approved Task creation and synthetic-data evaluations.
Structured LLM extraction, rule retrieval and a local FHIR sandbox, with human-approved Task creation and synthetic-data evaluations.
The challenge
Referral information can be incomplete or ambiguous. An assistant must find eligible guidance, expose uncertainty and keep administrative checks and write permission in server code.
Product & engineering
A referral intake application where staff review extracted fields, missing information and rule evidence, then approve an exact FHIR Task draft. Evaluation uses synthetic referrals and administrative rules for a fictional clinic.
Structured LLM extraction with deterministic labeled-field verification, plus distinct Fixture and Connected modes.
Lexical, model-vector and hybrid rule retrieval, filtered by scope and effective date before ranking.
A local FHIR R4 sandbox for Patient, ServiceRequest and Task, with exact human approval and durable commit idempotency.
A deterministic baseline and bounded agent variant, with permitted read tools, run traces and step, call, token and time limits.
A read-only MCP search interface checked by an independent HTTP client, and a 40-case synthetic evaluation set with ten workflow controls.
System design
Synthetic notes and administrative fields.
Extraction, rule retrieval and coded checks.
A person inspects the exact Task payload.
Server approval checks and idempotent commit.
One Spring Boot service and Angular desk use PostgreSQL for rules, vectors, local FHIR resources, runs and committed Tasks. The baseline performs extraction, lookup, retrieval, coded checks and drafting. A bounded agent selects permitted read tools; the server owns Task preparation, approval validation and commit. MCP exposes rule search only.
Outcome & evidence
Delivered outcome
A completed intake workflow with recorded connected evaluations, an independent MCP client check, and approval/idempotency browser checks. Error analysis led to a deterministic extraction safeguard; the agent comparison supported keeping a simpler baseline.
23 September 2026 · 40 synthetic referrals · 22 dev / 18 holdout · corpus clinic-admin-2026.09.1 · dataset referrals-2026.09.2 · gpt-4o-mini, extract-v3 and agent-v3.
An earlier prompt invented a member ID after an instruction embedded in a referral. The final prompt was revised after that holdout case was inspected. Citation checks verify eligible source and snippet membership, not semantic support for every claim. A fresh independent holdout is needed for broader quality claims.
Synthetic data and a local FHIR subset only: no real EHR integration, clinical or coverage decisions, production authentication or durable recovery of interrupted runs. Recorded measurements describe this small regression set and were not rerun on the portfolio site.
Connected care
Enterprise Engagement
Architecture and technical leadership across multiple teams delivering digital healthcare.
Architecture and technical leadership across multiple teams delivering digital healthcare.
The challenge
A multi-team telehealth estate needs one architectural direction, a coherent data model and release coordination across squads.
Product & engineering
Technical architecture and delivery leadership for a multi-team digital healthcare platform.
Architecture and design artifacts.
Database design.
Team technical direction and release coordination.
System design
Patient and clinician product surfaces.
High- and low-level system design.
A coherent database model across teams.
Client collaboration and release coordination.
Multi-team delivery under a shared high-level and low-level design, rather than isolated squads.
Outcome & evidence
Delivered outcome
Delivered as a live enterprise telehealth platform.
The client identity and internal implementation details are omitted.
Research workflow
Independent Product
Public market signals turned into systematic investment research, strategies and alerts.
Public market signals turned into systematic investment research, strategies and alerts.
The challenge
Alternative-data signals usually live in filings, vendor feeds and private notebooks. Readers need a place to inspect strategy families and unlock deeper history without opening a trading account.
Product & engineering
An independent research product built around public disclosures, systematic strategies and a subscription experience. Research only — not a brokerage and not personalized advice.
Public site: landing, pricing, FAQ, methodology, legal, news and shareable historical cases.
Strategy catalog for published families, curated picks and watches.
Subscription checkout with pluggable payment providers.
Email and Telegram alerts, including preference and stop flows.
Admin surfaces for jobs, data sources, ingestion and editorial publishing.
System design
Filings and alternative-data inputs.
Python analysis and strategy processing.
Java / Angular catalog and subscriptions.
Email and Telegram with user preferences.
Modular monolith: Spring Boot, Angular, PostgreSQL, Flyway and JWT, plus a Python research engine started as a job process. Access is subscription-based. Payment providers are feature-flagged so checkout can change without rewriting the research domain.
Outcome & evidence
Delivered outcome
The product is live at investivoq.com.
Platform modernization
Enterprise Engagement
Designing a phased transition from a legacy estate to cloud services and domain-oriented data.
Designing a phased transition from a legacy estate to cloud services and domain-oriented data.
The challenge
A domain-heavy legacy stack cannot cut over in one step. Data and writes have to move without stopping the business.
Product & engineering
Architecture for a phased migration from a legacy application and database estate to a cloud platform.
Target-cloud and data design.
Domain boundaries and coexistence of legacy and new systems.
Shadow mode, transition bridge, outbox and dead-letter handling.
System design
Existing applications, data and writes.
Domain boundaries, bridge and shadow mode.
Java services and AWS data architecture.
Outbox and dead-letter handling.
Strangler-style coexistence rather than a single big-bang rewrite. Java, PostgreSQL and cloud delivery, with domain-oriented data boundaries.
Outcome & evidence
Delivered outcome
Migration architecture and integration patterns are defined.
Internal programme and client names are omitted. The case focuses on migration architecture.
Assessment protocol
Independent Product
A local tool that validates three manually imported AI assessments and turns them into an inspectable fluency profile.
A local tool that validates three manually imported AI assessments and turns them into an inspectable fluency profile.
The challenge
Most “AI skill” scores count activity. That measures usage, not effectiveness, and it rewards the wrong evidence.
Product & engineering
An application for turning an assistant’s observations of real collaboration into a private, inspectable AI Fluency profile. Assessments are pasted in by hand, validated, stored and aggregated by explicit rules.
A versioned response format and manual import. The server checks the methodology version instead of trusting the pasted label.
Ten AI Fluency dimensions, with INSUFFICIENT_EVIDENCE kept as “no score” rather than a zero.
Immutable assessment history, methodology versions and a hash of the stored payload.
A privacy guard that drops known raw identifiers. It is a filter, not a guarantee.
Two aggregation rules: a median across sources for AI Fluency, and the strongest-evidence report for a technology claim.
System design
Assistant context from real collaboration.
Versioned response and server validation.
Immutable assessments and source records.
Explicit aggregation and evidence review.
The working context stays with the assistant. A structured response is pasted in, checked, and stored as a versioned assessment. AI Fluency uses the latest compatible report from each source and takes the median. A technology level is not averaged: the report with the highest evidence strength wins, then confidence, then the later date.
Inside the application

Usage counts do not show the quality of human–AI collaboration.
Protocol, validation, versioned storage and an inspectable review interface.
Three assessments in a browsable local profile with traceable rules.
There is no automatic link to ChatGPT or Cursor. A report only covers the context that assistant could see, and the owner pastes it in by hand.
Real profile snapshot · Methodology ai-fluency-2.3
Assessments reflect each assistant’s available context. They are not an independent skills certification.
AI Fluency · median across three sources
8.6Range across sources: 7.8–9.1Latest compatible result from each source. Sorted values: 7.8, 8.6, 9.1 → middle value = 8.6. Another source does not add points.
Each figure is the median of the sources that scored that dimension.
Technology claims follow a different rule: evidence strength → confidence → later date.
The two 5/5 reports remain. Confidence 0.95 wins over 0.88, so the selected level is 8.4.
The 4.5 report is not averaged in. It has lower evidence strength for this claim.
Observer labels are kept as written. Amazon Web Services and Amazon Web Services (AWS) stay separate, and so do the two AI agent orchestration names.
A summary of the observer reports, not facts established by an independent audit.
Spring Boot, Angular, PostgreSQL, Flyway. Immutable history, versioned rules, a hash of the stored payload, and a privacy guard that drops known raw identifiers. The guard is a filter, not a guarantee.
Outcome & evidence
Delivered outcome
Three manually imported assessments are on the local profile. The AI Fluency median is 8.6, with a source range of 7.8–9.1. This is an inspectable local record, not an independent certification.
An AI report only covers the context that assistant could see. Evidence and confidence are self-reported by the observer. Agreement between sources is not an external exam, and a technology score is not a substitute for code, experience or an interview.