[ Start ]
A name or a topic
Farming starts with a customer name, single or as a bulk import from the base. Hunting starts with topic, description, and region. No more input is needed; the agents do the rest.
[ deep research × sales ]
Deep research for sales: farming and hunting.
I built this tool for my own use: to support the go-to-market of AI services in a consulting organisation. Two modes share the same research engine. Farming: a customer name from the existing base goes into the pipeline, and deep-research agents work through profile, challenges, and prioritised AI use cases in sequence. Hunting: a search profile of topic and region becomes a validated target list, every verdict with a source. This page shows how the system is built, which decisions stand behind it, and what was learned along the way.
Agent chain
The agents run in sequence and build on each other: the profile feeds the challenges, both feed the use cases.
Farming
A customer name is enough: the outcome is sourced dossiers, prioritised AI use cases, and a competitor benchmark.
Hunting
Topic and region become a validated target list: breadth search first, then deep validation with a source per verdict.
Costs under control
Every run is accounted to the cent: budgets per agent, quotas per user, no double billing.
[ preparation beats cold calling ]
The best pitch starts long before the conversation.
Selling AI services rarely fails on the offering; it fails on the opening: generic pitches reach nobody, and individual preparation costs hours per company. Deep research makes that preparation scalable: agents with live access to the web do the research that used to be manual work and deliver sourced talking points instead of guesses.
The agents are no black box. Each agent is an editable research brief: the team can rephrase it, reorder it, or add its own, without code. And they work as a chain rather than one-shot: every agent receives its predecessors' results as context, and the order determines what later agents build on.
And because every research run costs real money, cost discipline is part of the architecture: budgets per agent, quotas per user, accounting per run, no double billing. Research at scale only works when the bill adds up.
[ two modes, one engine ]
From a name to a dossier, from a topic to a list
Farming and hunting share the same research engine and differ only in input and output. Every run passes through these stations.
[ Research ]
Agents with live access
Deep-research agents search the open web: annual reports, press, job postings, industry directories, trade-fair lists. Progress is visible live, to the minute, per agent.
[ Build ]
The chain, not one shot
In farming, three agents interlock: profile and process landscape, then challenges, then prioritised AI use cases with a competitor benchmark. Each receives its predecessors' results, and the order is set by drag and drop.
[ Condense ]
From text to tools
From the dossiers, one click produces a slide deck per use case, a personal email draft, and a newsletter as an Outlook file. Plus a chat assistant that stays strictly inside the research and names its gaps honestly.
[ Validate ]
Hunting with a citation rule
First the breadth search in several passes, until nothing new appears. Then deep validation in small batches: confirmed, unconfirmed, or rejected, every verdict with a source. Nothing is deleted; even rejections stay traceable.
[ Hand over ]
Into daily sales work
The target list goes to sales as an Excel file, CRM contacts are connected, and the dossiers provide the basis for the conversation. Research becomes a well-prepared first meeting.
[ architecture ]
Three layers, one research brief
Research engine
Deep research with live web search, no crawler of its own
The actual research is done by a hosted deep-research model with native web search: the system phrases the brief, the model searches the web on its own and returns a report plus a source list. No crawler, no scraping, no search-engine plumbing to maintain.
Underneath runs a non-blocking state machine: a waiting agent occupies no compute slot, and each run's state lives in the database and survives restarts. Runs may take hours, and stranded jobs can be resumed precisely, without paying again for agents that already succeeded.
Partial results count too: if a run hits the output limit, a tolerant parser salvages the complete rows instead of discarding everything. Half a market scan is worth more than none.
Agents & knowledge
Editable research briefs, chained per project
Agents are editable briefs in the database, not code: three defaults per farming project, two per hunt. The team rephrases them, reorders them by drag and drop, and adds its own; the chain is as long as desired.
Results land in three usable forms: as a dossier with sources, as searchable full text across all customers, and as a vector per customer and focus for similarity search. Topic clusters across the whole base emerge automatically on top.
In hunting, a multi-stage dedupe protects the list: the website domain beats the name, legal forms are normalised, and filled fields are never overwritten by empty ones.
Sales layer
Dossiers, slides, drafts, lists, cockpit
Research becomes tools: a slide deck per use case, an email draft streamed live, a newsletter as an Outlook file with the image embedded. The chat assistant stays strictly grounded in the research and says honestly when information is missing.
The target list shows its verification states openly: confirmed, unconfirmed, rejected, with reason and source per row, exportable as an Excel file. CRM contacts are read and matched.
And costs remain part of the layer: quotas are checked before a run starts, a priority queue lets interactive research overtake bulk imports, and every run is billed individually.
[ lessons ]
Research that costs money needs operational discipline.
The formative incident: the model provider's credit ran out, runs failed, and started research could not be resumed, only deleted and paid for again. That became the foundation of the architecture: durable job state in the database, resumption exactly where a run stopped, and no double billing, even if a process crashes mid-run.
Hunting was just as instructive: in the first version, every agent result replaced the entire list, and whatever the last agent failed to repeat was lost. Today results are merged instead of replaced, truncated output is repaired instead of discarded, and the search runs in passes until a market is truly exhausted. Loss-free handling is not a comfort feature in paid research; it is a duty.
The platform came together solo in four and a half months: 335 commits, around 35,000 lines of application code and almost 11,000 lines of tests, every major step preceded by a design document. Built for the step from a handful to hundreds of users.
[ numbers ]
The system in numbers
3 + 2
deep-research agents in the default chain: three for farming, two for hunting, all freely editable
50+
companies from a typical market scan in hunting, searched in passes until nothing new appears
5
tools per customer from one research run: dossier, slide deck, email draft, newsletter, and chat assistant
to the cent
every research run is accounted for: quotas per user, checked before the start, no double billing
[ technology ]
Technology in use
OpenAI
FastAPI
Celery
Redis
PostgreSQL
OpenAI
FastAPI
Celery
Redis
PostgreSQL
Elasticsearch
React
TypeScript
Docker
Azure
Elasticsearch
React
TypeScript
Docker
Azure
Put the potential of this technology to work.
Wondering how deep research could support your market development or sales? Then let's talk: you bring the challenge, I bring the experience and the tools.