How I Work - Amar Harolikar

The work

Two kinds, same projects. Analytics, full-stack apps, the infra they run on. Built for small and medium businesses, not enterprises.

Sessions

Hands-on, over screen share. I work through your actual problem with you and get you up to speed on the tool .. Claude Code in VS Code, the Claude app, Claude in Excel and PowerPoint. What I cover is your call: frontend, backend, databases, models, pulling data from different sources, a full analysis end to end.

Two kinds of folks come for sessions. Experienced practitioners (typically 10-15 years in) who want to use the tool on their own work. And division heads .. CEOs, CFOs, heads of marketing, analytics, operations, 20 to 30+ years in .. who want the architecture and the flow at a higher level, plus how to do their own role better with it.

Either way the sessions still center on you actually using the tool .. that is what makes the time pay off. You want to use it for your own work, and understand how your team works with it.

Every session is one to one. Not a group call or a training batch - your own problem & questions, your own pace. All data and discussions stay private.

Execution

I build it. Database AI, report generators, machine learning models, automations, dashboards, full-stack apps with auth and access control. The heavy lifting and the tedious pieces that take time to get right. Same projects as the sessions .. I just do the work.

What I take on

Where I go deep. Analytics and machine learning, AI tools and agentic apps, full-stack builds, self-hosted infrastructure, production security, the Cloudflare edge, agent discoverability and basic SEO. Live examples under each one.

Many examples of all of this are live on the main site .. apps, code, posts. Most are open source. Browse /browse-apps, /analysis, /blog.

Analytics and machine learning models

Large part of what I do .. analysis, dashboards, ad-hoc reports, machine learning models. Data science and decision sciences. As standalone offline projects or as full-stack interactive tools connected to live systems.

Live examples

TREMOR (multi-source macro signals across 8 categories), MFPRO (mutual fund analytics with composition drift), IMDB Dashboard (12M+ titles, 230M-row backend), VIGIL (Indian markets + corporate signals across 7 NSE tables). All listed at /browse-apps.

AI tools and agentic apps

A specialty. Helping figure out where AI fits and in what form .. sometimes a single AI app is enough, sometimes API chaining, sometimes a full agentic setup. Then everything that sits on top of the build .. orchestration, observability, validation, hallucination controls, and the cost-management layer (per-user budgets, model routing) that matters as much as the work itself .. smart models are excellent but expensive. Live AI apps in use by medium-sized teams, mostly for operations.

See it run

The Database AI suite at /database-landing .. multi-step reasoning agents that connect to multi-GB databases, write SQL, run analysis, generate PDF reports. MCP server directory at /mcp-landing is callable from any AI client (Claude, Cursor, custom).

Full-stack apps

Mostly React + Vite frontends, FastAPI backends, Postgres or DuckDB. Platform-agnostic though .. PHP / Apache / MySQL and other stacks too. Pick up whatever the team already runs on.

Live examples

TREMOR, MFPRO, IMDB Dashboard and VIGIL are full-stack apps shipped without login on purpose .. they're free public tools, no friction. DuckIt is the auth example .. guest sessions get 2-day file storage, logged-in users 7 days, owner level keeps files permanently. All on /browse-apps.

Self-hosted infrastructure

Hetzner and OCI VPS, Coolify for deploys, Docker, and Postgres self-hosted on the same setup. I moved most databases off managed Postgres .. the usage charges added up and self-hosting turned out simpler to run. Managed Postgres still gets used here and there.

What's running on it

Every app on /browse-apps and every API on /apis runs on this infra. tremor.tigzig.com, vigil.tigzig.com, mfpro.tigzig.com, briq.tigzig.com .. each its own subdomain, container and deploy.

Production security

Hardening applied across the whole stack .. cloud, frontend, backend, plus continuous monitoring of every request. A 120-item baseline checklist is published at /security as a useful starting point .. it's a sample of what I actually run.

The evidence

All my data APIs run open, no auth, public internet (see /apis) .. they take real-world attack traffic, and the hardening at /security is what came out of that. My ops command center sits in the open too .. logs.tigzig.com, multi-layer defenses. The app itself has delete access to ~200 GB of data and admin to 40+ apps and services, running in the open since launch.

Cloudflare edge

Cloudflare Workers in production. Honeypot routes catch scanners; offenders go into a jail and get blocked at the Cloudflare WAF. Plus rate limits, agent-traffic logging, markdown content negotiation and subdomain SEO.

Where you see it

Every page on tigzig.com is fronted by these Workers .. every frontend and backend. The agent-traffic logger feeds an ops dashboard.

AI-agent discoverability (AEO)

tigzig.com is built agent-first. RFC 9727 catalog at /apis, /llms.txt with an intent map, MCP server directory at /mcp-landing, markdown twins of every page. Point an AI agent at the site and it maps the whole platform in one fetch.

The check

Ask any AI agent .. ChatGPT, Claude, Perplexity .. "what is tigzig.com". It will map the site from /llms.txt within seconds .. apps, APIs, MCP servers, recent posts. Including content published this week.

Basic SEO

Basic mechanics to keep pages findable .. sitemap structure, JSON-LD breadcrumbs, structured data, Bing Webmaster + GSC + IndexNow integrations. The plumbing underneath the AEO work above .. not deep SEO consulting.

The check

Tigzig pages show up in Google, Bing and ChatGPT's live web search. Nothing fancy, just the basics keeping pages findable.

Open-source analytical tools

Free tools I've built and shared. TREMOR for global macro signals, VIGIL for Indian markets and company-level info, MFPRO for mutual fund analytics, plus a set of portfolio and technical analysis tools, database converters and text-to-SQL apps. The data behind them is on /apis as zero-auth APIs and MCP servers.

How execution works

Billed on brain time rather than the clock, every sprint pre-approved with an hour cap, the repo stays in your hands, and it goes live early.

Billed on brain time, not the clock

I charge for the time I am actually on your project .. planning, reviewing, validating, verifying output. Claude Code does the heavy lifting on its own. I might be at my desk 10 hours running three projects, you get charged for the hours on yours, not the wall clock.

Calls during an execution project .. planning, reviews, screen-share walkthroughs .. are billed on call time, at the execution rate.

Every sprint is pre-approved

Before a sprint I tell you roughly what I am planning. You approve an upper limit, say 3 to 5 hours or 5 to 10. I stop at that limit and report back. No runaway bills, no surprises. The plan is the intent .. what I aim to do within the cap. You see both up front.

The repo is yours. You can take over any time

Repo and documentation sit with you, or you have full control. Stop the project, pick it up later, hand it back .. no issue either way. Most of my clients work with Claude Code too, so the handoff stays clean. Same docs, same tool.

It goes live early, and I stay on it

These are not month-long builds handed over at the end. Your real users, usually a CXO's team, are on the app from 30 to 40% in, the moment basic flows work. So by the time we reach 90%, most of their issues are already fixed and it is in real use .. days to weeks, not months. By the final stretch, most clients prefer to handle the small UI tweaks themselves .. faster and cheaper at that point, and they know the app well by then. Flexible either way, they can do it or I can. I keep supporting through full rollout and beyond, stepping in whenever you want.

Rates

SessionsGetting up and running, setup help, build guidance

₹9,600 / hr

ExecutionBuild work, charged on brain time

₹2,400 / hr

Rates vary by region and engagement. The above are indicative.

All sessions and execution are prepaid .. PayPal, Razorpay, UPI, or bank transfer, whichever works for you.

Ongoing work. The work can also run as a monthly retainer, starting at ₹24,000. That is a minimum rather than a fee on top .. it is set off against actual work at a reduced rate, so what you pay for is the work, with the monthly figure as the floor. It reserves a block of my time each month, and it puts you ahead of one-off work when something new comes up. Most retainers start after a first project or two, once we both know how the work runs.

It also means I keep an eye out between projects. The AI companies change something every week and most of it is noise for any one business, so when a piece of it is actually relevant to you, you get a note without having to ask. That could be a model that would cut your API bill, something one of the frontier labs has released that fits your sector, a security advisory on the stack you run, or a tool someone else in your space has built and what it does.

How many hours do I need?

Sessions. A first one usually runs around 2 hours. Setup and installation eat into that if it is not already done. After that it genuinely varies, with your background and with the problem. Someone coming from analytics, from programming, or from marketing all start at a different place, and so does someone already working in VS Code or a CLI tool versus someone who has never opened one. Some things get resolved in a single sitting, some run across several sessions.

Execution. Scope-based. We work out what you want built, I come back with an estimate in hours, and every sprint is capped and approved before it starts. Some examples from real engagements are below.

Some examples

Hours from real engagements, all anonymised, from 90 minutes up to about 70 hours. These are brain-time hours .. the calendar time is longer, and is shown next to each one. The rate is right above, so you can work out what that comes to.

Execution

Paid scoping prototype. Fully clickable, dummy data, built to make a proposed system concrete before anyone commits to it.

~2 hrs

Operational monitoring dashboard. Live operational parameters and activity tracking, pulled together into one view the team can actually scan.

~15 hrsbrain time, over about a week

Response model, delivered as a tool. The model itself was built offline. The tool is what the team runs afterwards: pull the latest data from their sources, score it, and put the new run side by side with the previous one, so a shift in the population splits or the decile table shows up on its own rather than being noticed months later. UI on top, Python underneath, runs locally.

~25 hrsbrain time, over about 2.5 weeks

Full-stack AI-based tool built into a client's existing stack. Records from their database go out to an AI for review and correction, and the results are written back after checks. Wired to their existing auth with role-based access, every API call and database update logged, observability at each stage. They took it over at around the 90% mark and finished the tweaks themselves.

~70 hrsbrain time, over 4 to 6 weeks

Sessions

A CEO, on Claude Cowork and Claude in Excel and PowerPoint. Reviewing the work coming up from his juniors, modifying decks, quick prototypes. A getting up and running session, then a deep dive later on specific problems.

90 minplus 1 hr later

A head of analytics, around 20 years in. A getting up and running session, demo plus hands-on, building machine learning and statistical models and automations with Claude Code.

~2.5 hrs

A CFO, 25+ years in finance and new to all of this. Worked through the basics .. what an API is, where model cost sits, subscription versus API cost, build versus buy, model routing, and how an agentic setup works. All shown live on screen.

~5 hrsover 2 to 3 sessions

A founder coding through a chat AI, moved to Claude Code. Deep-dive sessions where I walked him through setting up the automated deploy path out to his cloud server, live on screen. He went on to refactor his own sites himself, and later used the same workflow to rebuild his setup after a security incident on his server.

8 to 10 hrsseveral 2 to 3 hr sessions

These are indicative, not a scale. Each number belongs to that one piece of work, and what any other piece takes depends entirely on its scope. The spread runs wide in both directions. I have done scoping work and dummy-data prototypes that ran a lot longer than the two hours shown above, and I have built small full-stack apps in far less time than the larger ones here. Two projects that read alike on paper can land a long way apart once the detail turns up. It is a rough sense of the ground rather than a yardstick to measure your own work against. The real number comes out of scoping, and every sprint is capped and approved before it starts.

Best used with an AI agent

Sessions and execution. Same work, seen two ways.