Writing

176 engineering articles, worked applications and research notes. Search the library, or pick a series.

Building a solo company with AI agents

40 articles · 4.3 h

An experimental operating platform for one founder: coordinate enquiries, sales, delivery, support and finance through shared records, controlled actions and an approval inbox.

01A solo company is a set of promises that must be kept8 min 02The company needs a memory that can answer “what changed?”8 min 03An agent may propose a decision; an action decides whether it can happen7 min 04The harness coordinates the work; it does not become the company8 min 05Connectors translate the outside world into facts and receipts10 min 06Market to lead: turn an enquiry into a decision worth making7 min 07Campaigns and outbound: keep the send button separate from the draft7 min 08Lead to cash: make the commercial commitment explicit7 min 09Invoice to cash: a payment is a fact, not a reassuring message7 min 10Retention and renewal: connect the support history to the next promise7 min 11Order to delivery: complete the promise, not just the task list7 min 12Issue to resolution: make “resolved” mean the customer can move on7 min 13Procure to pay: an approved need is not an approved payment7 min 14Hire to productivity: treat capacity as more than a headcount7 min 15Record to report: make the operating picture explainable7 min 16Risk to control: the same rule must apply in every loop7 min 17Concept to launch: help the founder decide what deserves building7 min 18Plan to performance: a dashboard should lead to a recorded decision7 min 19Operate to improve: an agent should not grant itself more authority7 min 20One customer journey across the loops8 min

21What does it cost to delegate a business task?7 min 22Evaluate the business state, not just the answer7 min 23The founder’s cockpit and the deployment behind it7 min 24Build one complete operating loop, then widen it8 min 25Durable workers: the company must survive a process restart5 min 26Memory that supports a decision without rewriting the company5 min 27Connect the mailbox and CRM without losing ownership of the facts5 min 28Documents, calendars and code: connectors for delivering the promise5 min 29Payment events, accounting entries and the meaning of “paid”5 min 30A capability registry for tools the agent may actually use5 min 31Parallel agents without conflicting commitments5 min 32When the customer document tries to become the system prompt5 min 33Trace the business task, then debug the agent5 min 34A small deployment that can grow into a dependable platform5 min 35Build a benchmark around the company’s promises5 min 36Polsia and the systems around the autonomous-company idea5 min 37What agent research changes in this business platform5 min 38Three businesses you could model with these loops5 min 39Run the company lab: a quote, a lost receipt and a partial payment5 min 40From shadow mode to a supervised operating loop5 min

Series overview

Building agentic patient workflows

18 articles · 1.8 h

An implementation series on patient operations: domain agents, action cards, approvals, FHIR and messaging connectors, durable workflows, evaluation and deployment.

01Start with the patient workflow7 min 02From language to domain actions8 min 03Making intent visible in the interface7 min 04Confirmation is part of the workflow7 min 05Patient journeys beyond the chat8 min 06Context, memory and live updates7 min 07Patient identity and FHIR connectors6 min 08Booking under concurrency and time zones5 min 09Messaging connectors and delivery receipts5 min

10MCP, A2A and the tool boundary5 min 11A durable harness for patient work5 min 12Security at every agent boundary5 min 13Retrieval for policies and patient context5 min 14Voice, multilingual input and human handoff5 min 15Evaluating complete patient workflows6 min 16Operating the platform: capacity, cost and recovery5 min 17What Polsia and agent research change6 min 18Run the lab and ship one workflow5 min

Series overview

Engineering private financial AI

12 articles · 1.4 h

On-premises SLMs for wealth management: fine-tuning, Hindi and Hinglish tools, hardware, evaluations and a custom benchmark.

01Private financial AI: from a question to a verified answer7 min 02Financial documents to source-linked training data6 min 03Generate financial QA with executable answer checks7 min 04Fine-tuning a local financial assistant with QLoRA7 min 05Chat templates, loss masks and response contracts7 min 06English, Hindi and Hinglish tool use6 min

07A bounded runtime for financial agents7 min 08On-premises serving: memory, latency and capacity7 min 09Typed scoring for financial answers and agent traces7 min 10Benchmark protocols for FinQA, FinanceBench and IndiaFinBench7 min 11WealthOps: an executable financial-agent benchmark7 min 12From benchmark runs to a reproducible release7 min

Series overview

Engineering agent harnesses

18 articles · 2 h

A hospitality AI concierge and operations platform: dynamic skills, scoped tools, durable state, evaluation, observability and cost.

01From an AI concierge to an operations system8 min 02Designing the harness: architecture, Hermes and OpenClaw7 min 03The run contract: orchestration, lifecycle and idempotency7 min 04Skills and on-the-fly composition7 min 05Prompt assembly is an engineering pipeline7 min 06Memory: what was said, what was shown and what is true7 min 07The model transport and the bounded agent loop6 min 08Tools as data, execution as a controlled boundary6 min 09Authority, approvals and the trust boundary7 min

10Booking: let the model converse and the system decide6 min 11From a complaint to work that someone owns6 min 12One event spine, several ways to deliver the result6 min 13The control plane: changing behaviour without losing control6 min 14Evaluation starts with the operational scenario6 min 15Judges, human review and regression evidence6 min 16The operator console is part of the agent product7 min 17Latency, concurrency and the cost of a completed task7 min 18Reliability, release discipline and operational value7 min

Series overview

Engineering financial RAG

14 articles · 1.6 h

An equity analyst platform over a million financial documents: parsing economics, hybrid retrieval, paragraph citations and evaluation.

01A million documents, six weeks, and a $50,000 constraint7 min 02A financial corpus begins with identity, time and provenance7 min 03A correctly read number can still be the wrong financial fact7 min 04Bringing a $120,000 parsing iteration inside a $50,000 budget8 min 05Chunk the evidence, not just the text7 min 06Choosing embeddings for financial evidence6 min 07Finding the right issuer, period and passage6 min

08From fifty candidates to an answerable evidence set7 min 09A citation should take the analyst to the evidence7 min 10Make the calculation inspectable, too6 min 11When financial research needs more than one retrieval step7 min 12Evaluate the evidence chain, not just the final answer7 min 13Operating RAG across a million financial documents9 min 14The engineering decision that made the product affordable7 min

Series overview

Engineering conversational AI

12 articles · 2.6 h

Telephony voice agents: Bolna and Pipecat architecture, latency and turn-taking, multilingual speech, cloud deployment and evaluation.

01Conversational AI Design: the call is the product12 min 02Why an 800 ms voice response is a systems problem11 min 03A Bolna-based telephony agent, from connection to hangup10 min 04Designing a configurable voice platform with Pipecat10 min 05Choosing the speech and reasoning models for a voice agent10 min 06Self-hosted Whisper: streaming, scheduling and the concurrency trap10 min

07Adapting a 2026 speech recognizer for English, Hindi, Singlish and Hinglish13 min 08Speech synthesis that can survive a real customer call16 min 09Context, memory and tools: where a voice agent becomes a business system12 min 10Running voice agents on AWS and GCP without losing the conversation20 min 11Evaluating a voice agent: listen to the call, inspect the action19 min 12What a voice agent changes for the business11 min

Series overview

Engineering conversational avatars

8 articles · 0.7 h

Avashi's client-commissioned research: avatar generation, procedural animation, text and voice, persistent memory, and the mobile and budget constraints that closed the engagement.

01An internal avatar platform: the research brief and the approach map5 min 02Parametric generation: FLAME, Anny and the topology contract5 min 03Blender, imported avatars and the gap after generative 3D5 min 04Expressions, lip sync and the alternative Live2D path5 min

05Text and voice on top of the avatar runtime5 min 06Memory across conversations: identity, retrieval and correction5 min 07Why the rendered avatar was too heavy for mobile5 min 08Evaluating the complete avatar product—and knowing when to stop5 min

Series overview

Engineering creative generation pipelines

6 articles · 0.5 h

XGen's advertising and marketing workflows: recovered ComfyUI components, Nano Banana images, Wan 2.1 video, LoRA provenance, Runpod workers and creative evaluation.

01From a marketing brief to a reproducible creative pipeline5 min 02Image generation: Nano Banana, local models and structured prompts5 min 03Wan 2.1 image-to-video: an inspectable ComfyUI graph5 min

04LoRAs, model identity and the reproducibility record5 min 05Runpod workers: GPU memory, CPU work and job economics5 min 06Evaluate the creative asset, then release the pipeline5 min

Series overview